Merge branch 'master' of https://git.d-popov.com/popov/neuron-tai

2026-07-01 14:45:13 +03:00
parent 85a8496985 50ec507c7a
commit c5c60cf289
14 changed files with 684 additions and 35 deletions
--- a/QUICKSTART.md
+++ b/QUICKSTART.md
@@ -86,12 +86,18 @@ python -c "import torch; print(torch.__version__, torch.cuda.is_available())"

 If you get `ModuleNotFoundError: No module named 'torch'` even though `pip install torch`
 says "already satisfied", the `torch/` package directory is missing while the metadata
-stub remains (can happen after a conda-managed install). Force-reinstall via pip:
+stub remains (can happen after a conda-managed install). Force-reinstall all three
+PyTorch packages together so their versions stay in sync:

 ```powershell
-pip install --force-reinstall torch --index-url https://download.pytorch.org/whl/cu118
+pip install --force-reinstall torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
 ```

+> **Important:** always reinstall `torch`, `torchvision`, and `torchaudio` as a group.
+> Upgrading only `torch` leaves `torchvision` on an incompatible version, which causes
+> `RuntimeError: operator torchvision::nms does not exist` and makes transformers fail
+> to import any model class (the error surfaces as `Could not import module 'Qwen2ForCausalLM'`).
+
 Then re-run the verify step above.

 If that prints `True` but `meshnet-node` still can't find torch, the venv entry point
--- a/docs/issues/30-manual-route-and-hop-benchmark.md
+++ b/docs/issues/30-manual-route-and-hop-benchmark.md
@@ -0,0 +1,48 @@
+# US-020 — Manual route selection + hop-penalty benchmarking
+
+## Context
+
+The tracker auto-selects inference routes based on synthetic benchmark scores. To measure
+the real cost of adding hops (latency per node boundary), we need:
+
+1. A way to pin a request to a specific route so we control the variable.
+2. A benchmark endpoint that runs the same prompt through 1-node, 2-node, and 3-node
+   routes and records per-hop latency.
+
+Results are stored to disk. Routing algorithm is **not** changed in this story — this is
+data collection only. The data will inform a future routing optimisation story.
+
+## Design decisions (grilled 2026-07-01)
+
+| Decision | Choice |
+|---|---|
+| Route spec | Optional `route` field in JSON request body (list of node IDs) |
+| Trigger | Explicit only — `POST /v1/benchmark/hop-penalty` endpoint |
+| Auth | Header-presence stub (`Authorization` must be non-empty); real auth in future story |
+| Routing integration | Store data only; routing algorithm unchanged |
+| Persistence | Append to `benchmark_results.json` in tracker working dir; in-memory queryable |
+
+## Acceptance criteria
+
+- `POST /v1/chat/completions` accepts optional `"route": ["<node_id>", ...]` in the
+  request body. If present, the tracker uses those nodes in order instead of auto-selecting.
+  If absent, existing routing is unchanged (no breaking change for unaware clients).
+- Missing or invalid node IDs in `route` return HTTP 400 with a descriptive error.
+- `POST /v1/benchmark/hop-penalty` is auth-gated: requests without a non-empty
+  `Authorization` header return HTTP 401. Body: `{"model": "...", "prompt": "...",
+  "max_new_tokens": 64}`.
+- Benchmark fans out to up to three routes: 1-node (single node covering all layers),
+  2-node (two consecutive shard nodes), 3-node (three nodes) — using whatever is
+  currently registered. Routes with insufficient coverage are skipped, not errored.
+- Response includes per-route breakdown: `total_ms`, `per_hop_ms: [...]`,
+  `tokens_generated`, `route: [node_id, ...]`.
+- Results are appended to `<tracker_working_dir>/benchmark_results.json` (created if
+  absent) as a JSON array. Each entry includes timestamp, model, prompt hash, and the
+  per-route breakdown.
+- `GET /v1/benchmark/results` returns the stored results array. Also auth-gated.
+- Clients that never send `route` or call `/v1/benchmark/*` are completely unaffected.
+- Integration test: send the same prompt via a pinned 1-node route and a pinned 2-node
+  route; assert 2-node result has 2 entries in `per_hop_ms`; assert both records appear
+  in `benchmark_results.json`.
+- `python -m pytest` passes from repo root.
+- Commit only this story's changes.
--- a/docs/prd.json
+++ b/docs/prd.json
@@ -684,10 +684,36 @@
        "US-022"
      ],
      "completionNotes": "_relay_hop() added to torch_server.py. _get_remaining_route returns list[dict]. relay_bridge.py updated with body_base64 support. Tracker injects relay_addr into downstream hop dicts. 157 tests pass."
+    },
+    {
+      "id": "US-030",
+      "title": "30 — Manual route selection + hop-penalty benchmarking",
+      "description": "Two additions to the tracker. (1) Callers can pin an explicit inference route by passing an optional \"route\": [\"<node_id>\", ...] field in the POST /v1/chat/completions body. The tracker uses those nodes in order instead of auto-selecting; clients that omit the field are unaffected. (2) A new privileged POST /v1/benchmark/hop-penalty endpoint runs the same prompt through up to three routes (1-node, 2-node, 3-node) using whatever nodes are registered, records per-hop latency, and appends results to benchmark_results.json in the tracker's working directory. The routing algorithm is not changed — this story is data collection only. Auth is a header-presence stub (non-empty Authorization header required for benchmark endpoints).",
+      "acceptanceCriteria": [
+        "POST /v1/chat/completions accepts optional \"route\": [node_id, ...] in the request body; if present, tracker routes to those nodes in order; if absent, existing auto-routing is unchanged",
+        "Missing or invalid node IDs in route return HTTP 400 with a descriptive error message",
+        "POST /v1/benchmark/hop-penalty requires a non-empty Authorization header; missing/empty returns HTTP 401",
+        "Benchmark body: {\"model\": \"...\", \"prompt\": \"...\", \"max_new_tokens\": 64 (optional)}",
+        "Benchmark fans out to up to 3 routes (1-node, 2-node, 3-node) using currently registered nodes; routes with insufficient coverage are skipped, not errored",
+        "Benchmark response includes per-route entries: {\"route\": [node_id, ...], \"total_ms\": float, \"per_hop_ms\": [float, ...], \"tokens_generated\": int}",
+        "Results appended to <tracker_cwd>/benchmark_results.json (created if absent) as a JSON array; each entry includes ISO timestamp, model, sha256 of prompt, and per-route breakdown",
+        "GET /v1/benchmark/results returns the stored results array; also requires non-empty Authorization header",
+        "Integration test: pin a 1-node route and a 2-node route for the same prompt; assert 2-node result has 2 per_hop_ms entries; assert both records appear in benchmark_results.json",
+        "Clients that never send route or call /v1/benchmark/* are completely unaffected (existing tests pass unchanged)",
+        "python -m pytest passes from repo root",
+        "Commit only this story's changes"
+      ],
+      "priority": 30,
+      "status": "open",
+      "notes": "Source issue: .scratch/distributed-inference-network/issues/30-manual-route-and-hop-benchmark.md. Design decisions grilled 2026-07-01: route via body field, explicit-only benchmark trigger, auth stub, routing algorithm unchanged, persist to JSON file.",
+      "dependsOn": [
+        "US-014",
+        "US-019"
+      ]
    }
  ],
  "metadata": {
-    "updatedAt": "2026-06-29T15:35:00.000Z",
+    "updatedAt": "2026-07-01T00:00:00.000Z",
    "statusVocabulary": {
      "open": "Not started",
      "in-design": "Decisions pending before implementation can begin",
--- a/packages/node/meshnet_node/model_backend.py
+++ b/packages/node/meshnet_node/model_backend.py
@@ -4,6 +4,7 @@ from __future__ import annotations

 import base64
 from dataclasses import dataclass
+from pathlib import Path
 from typing import Any, Literal

 Quantization = Literal["bfloat16", "int8", "nf4"]
@@ -65,6 +66,7 @@ class TorchModelShard:
        shard_start: int,
        shard_end: int,
        quantization: Quantization = "bfloat16",
+        cache_dir: Path | None = None,
    ) -> None:
        if shard_start < 0 or shard_end < 0 or shard_start > shard_end:
            raise ValueError("shard_start must be <= shard_end and non-negative")
@@ -89,9 +91,9 @@ class TorchModelShard:
                model_id,
                quantization_config=quant_config,
                device_map="auto" if quant_config is not None else None,
-                torch_dtype=torch.bfloat16,
+                dtype=torch.bfloat16,
                low_cpu_mem_usage=True,
-                use_safetensors=True,
+                cache_dir=str(cache_dir) if cache_dir is not None else None,
            )
            if quant_config is None:
                self.model.to(self.device)
@@ -104,7 +106,10 @@ class TorchModelShard:
            raise

        self.model.eval()
-        self.tokenizer = AutoTokenizer.from_pretrained(model_id)
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            model_id,
+            cache_dir=str(cache_dir) if cache_dir is not None else None,
+        )
        self.layers = _model_layers(self.model)
        self.total_layers = len(self.layers)
        # shard_end is INCLUSIVE (last layer index, 0-based), matching the CLI convention.
@@ -336,8 +341,9 @@ def load_torch_shard(
    shard_start: int,
    shard_end: int,
    quantization: Quantization = "bfloat16",
+    cache_dir: Path | None = None,
 ) -> TorchModelShard:
-    return TorchModelShard(model_id, shard_start, shard_end, quantization)
+    return TorchModelShard(model_id, shard_start, shard_end, quantization, cache_dir)


 def _model_layers(model: Any) -> Any:
--- a/packages/node/meshnet_node/model_catalog.py
+++ b/packages/node/meshnet_node/model_catalog.py
@@ -2,7 +2,10 @@

 from __future__ import annotations

+import json
 from dataclasses import dataclass
+from importlib.resources import files
+from pathlib import Path


@dataclass
@@ -15,6 +18,7 @@ class ModelPreset:
    vram_int8: float
    vram_bf16: float
    description: str
+    metadata: dict | None = None

    def vram_for_quant(self, quant: str) -> float:
        """Return VRAM requirement in GB for the given quantization."""
@@ -41,6 +45,25 @@ class ModelPreset:
        return None


+def _load_model_metadata() -> dict[str, dict]:
+    try:
+        raw = files("meshnet_node").joinpath("model_metadata.json").read_text()
+        data = json.loads(raw)
+    except Exception:
+        return {}
+    models = data.get("models", {})
+    if not isinstance(models, dict):
+        return {}
+    return {
+        str(repo): metadata
+        for repo, metadata in models.items()
+        if isinstance(metadata, dict)
+    }
+
+
+_MODEL_METADATA = _load_model_metadata()
+
+
 CURATED_MODELS: list[ModelPreset] = [
    ModelPreset(
        name="Qwen2.5-0.5B-Instruct",
@@ -123,6 +146,16 @@ CURATED_MODELS: list[ModelPreset] = [
        vram_bf16=16.0,
        description="DeepSeek's efficient MoE — strong coding + reasoning",
    ),
+    ModelPreset(
+        name="Kimi-K2.7-Code",
+        hf_repo="unsloth/Kimi-K2.7-Code",
+        num_layers=61,
+        vram_nf4=500.0,
+        vram_int8=1000.0,
+        vram_bf16=2000.0,
+        description="Large coding-focused MoE model",
+        metadata=_MODEL_METADATA.get("unsloth/Kimi-K2.7-Code"),
+    ),
 ]


@@ -140,6 +173,44 @@ def detect_num_layers(hf_repo: str) -> int | None:
        return None


+def model_metadata_for(
+    hf_repo: str,
+    num_layers: int | None = None,
+    cache_dir: Path | None = None,
+) -> dict:
+    """Return operator-facing model metadata for a HuggingFace repo."""
+    for model in CURATED_MODELS:
+        if model.hf_repo == hf_repo:
+            metadata = dict(model.metadata or {})
+            metadata.setdefault("num_layers", model.num_layers)
+            return metadata
+
+    metadata: dict = {}
+    if num_layers is not None:
+        metadata["num_layers"] = num_layers
+    try:
+        from transformers import AutoConfig  # type: ignore[import]
+
+        cfg = AutoConfig.from_pretrained(
+            hf_repo,
+            cache_dir=str(cache_dir) if cache_dir is not None else None,
+        )
+        for attr, key in (
+            ("model_type", "architecture"),
+            ("num_hidden_layers", "num_layers"),
+            ("hidden_size", "hidden_size"),
+            ("num_attention_heads", "attention_heads"),
+            ("vocab_size", "vocabulary_size"),
+            ("max_position_embeddings", "context_length"),
+        ):
+            value = getattr(cfg, attr, None)
+            if value is not None:
+                metadata[key] = value
+    except Exception:
+        pass
+    return metadata
+
+
 def browse_hf_hub(top_n: int = 20) -> list[dict]:
    """Fetch top downloaded text-generation models from HuggingFace Hub."""
    try:
--- a/packages/node/meshnet_node/model_metadata.json
+++ b/packages/node/meshnet_node/model_metadata.json
@@ -0,0 +1,32 @@
+{
+  "models": {
+    "unsloth/Kimi-K2.7-Code": {
+      "architecture": "Mixture-of-Experts (MoE)",
+      "total_parameters": "1T",
+      "activated_parameters": "32B",
+      "num_layers": 61,
+      "dense_layers": 1,
+      "attention_hidden_dimension": 7168,
+      "moe_hidden_dimension_per_expert": 2048,
+      "attention_heads": 64,
+      "experts": 384,
+      "selected_experts_per_token": 8,
+      "shared_experts": 1,
+      "vocabulary_size": 160000,
+      "context_length": 256000,
+      "attention_mechanism": "MLA",
+      "activation_function": "SwiGLU",
+      "vision_encoder": "MoonViT",
+      "vision_encoder_parameters": "400M",
+      "license": "modified-mit",
+      "native_quantization": "int4",
+      "download_size_gb": 595,
+      "recommended_short_name": "kimi-k2.7",
+      "recommended_engines": [
+        "vLLM",
+        "SGLang",
+        "KTransformers"
+      ]
+    }
+  }
+}
--- a/packages/node/meshnet_node/startup.py
+++ b/packages/node/meshnet_node/startup.py
@@ -15,6 +15,7 @@ from typing import Any

 from .downloader import compute_shard_checksum, download_shard
 from .hardware import detect_hardware, benchmark_throughput_checked
+from .model_catalog import model_metadata_for
 from .relay_bridge import RelayHttpBridge, peer_id_from_wallet
 from .server import StubNodeServer
 from .torch_server import TorchNodeServer
@@ -422,6 +423,9 @@ def run_startup(
        user_pinned_shard = shard_start is not None or shard_end is not None
        # Auto-detect shard range from model config if not explicitly provided
        if shard_start is None or shard_end is None:
+            try:
+                detected = _detect_num_layers(model_id, cache_dir=cache_dir)
+            except TypeError:
                detected = _detect_num_layers(model_id)
            if detected is None:
                raise ValueError(
@@ -459,6 +463,7 @@ def run_startup(
            quantization=quantization,
            tracker_url=tracker_url,
            route_timeout=route_timeout,
+            cache_dir=cache_dir,
            debug=debug,
        )
        _node_start_time = time.monotonic()
@@ -495,6 +500,7 @@ def run_startup(
            "score": 1.0,
            "tracker_mode": (shard_start == 0),
            "managed_assignment": not user_pinned_shard,
+            "model_metadata": model_metadata_for(model_id, total_layers, cache_dir=cache_dir),
            **registration_capabilities,
            **relay_fields,
        }
@@ -559,6 +565,7 @@ def run_startup(
            quantization=quantization,
            tracker_url=tracker_url,
            route_timeout=route_timeout,
+            cache_dir=cache_dir,
            debug=debug,
        )
        _node_start_time = time.monotonic()
@@ -587,6 +594,7 @@ def run_startup(
            "score": 1.0,
            "tracker_mode": (assigned_shard_start == 0),
            "managed_assignment": True,
+            "model_metadata": model_metadata_for(assigned_hf_repo, assigned_num_layers, cache_dir=cache_dir),
            **registration_capabilities,
            **relay_fields,
        }
@@ -722,11 +730,14 @@ def run_startup(
    return node


-def _detect_num_layers(model_id: str) -> int | None:
+def _detect_num_layers(model_id: str, cache_dir: Path | None = None) -> int | None:
    """Fetch num_hidden_layers from HuggingFace model config (downloads ~1 KB config.json only)."""
    try:
        from transformers import AutoConfig  # type: ignore[import]
-        cfg = AutoConfig.from_pretrained(model_id)
+        cfg = AutoConfig.from_pretrained(
+            model_id,
+            cache_dir=str(cache_dir) if cache_dir is not None else None,
+        )
        return int(cfg.num_hidden_layers)
    except Exception as exc:
        print(f"  Warning: could not read model config from HF: {exc}", flush=True)
--- a/packages/node/meshnet_node/torch_server.py
+++ b/packages/node/meshnet_node/torch_server.py
@@ -12,6 +12,7 @@ import urllib.error
 import urllib.parse
 import urllib.request
 import uuid
+from pathlib import Path
 from typing import Any

 from .model_backend import (
@@ -682,6 +683,7 @@ class TorchNodeServer:
        tracker_mode: bool | None = None,
        tracker_url: str | None = None,
        route_timeout: float = 30.0,
+        cache_dir: Path | None = None,
        debug: bool = False,
    ) -> None:
        self._host = host
@@ -691,11 +693,13 @@ class TorchNodeServer:
            shard_start,
            shard_end,
            quantization,
+            cache_dir,
        )
        # Auto-detect tracker mode: enabled when shard_start == 0 or explicitly set
        self._tracker_mode = tracker_mode if tracker_mode is not None else (shard_start == 0)
        self._tracker_url = tracker_url
        self._route_timeout = route_timeout
+        self._cache_dir = cache_dir
        self._debug = debug
        self._server: _TorchHTTPServer | None = None
        self._thread: threading.Thread | None = None
@@ -745,6 +749,9 @@ class TorchNodeServer:
            f"  [node] loading reassigned shard: {model_id} layers {shard_start}-{shard_end}",
            flush=True,
        )
+        try:
+            new_backend = _load_backend(model_id, shard_start, shard_end, quantization, self._cache_dir)
+        except TypeError:
            new_backend = _load_backend(model_id, shard_start, shard_end, quantization)
        self._backend = new_backend
        self._tracker_mode = shard_start == 0
@@ -797,12 +804,13 @@ def _load_backend(
    shard_start: int,
    shard_end: int,
    quantization: str,
+    cache_dir: Path | None = None,
 ) -> TorchModelShard:
    from .model_backend import load_torch_shard

    quant = validate_quantization(quantization)
    try:
-        return load_torch_shard(model_id, shard_start, shard_end, quant)
+        return load_torch_shard(model_id, shard_start, shard_end, quant, cache_dir)
    except MissingModelDependencyError:
        raise
    except InsufficientVRAMError as exc:
--- a/packages/node/pyproject.toml
+++ b/packages/node/pyproject.toml
@@ -27,3 +27,6 @@ meshnet-node = "meshnet_node.cli:main"
 [tool.setuptools.packages.find]
 where = ["."]
 include = ["meshnet_node*"]
+
+[tool.setuptools.package-data]
+meshnet_node = ["*.json"]
--- a/packages/tracker/meshnet_tracker/model_presets.json
+++ b/packages/tracker/meshnet_tracker/model_presets.json
@@ -0,0 +1,37 @@
+{
+  "models": {
+    "kimi-k2.7": {
+      "layers_start": 0,
+      "layers_end": 60,
+      "hf_repo": "unsloth/Kimi-K2.7-Code",
+      "aliases": [
+        "kimi-k2.7",
+        "Kimi-K2.7-Code",
+        "unsloth/Kimi-K2.7-Code"
+      ],
+      "recommended": true,
+      "deployment_status": "recommended",
+      "required_model_bytes": 638876385280,
+      "download_size_bytes": 638876385280,
+      "native_quantization": "int4",
+      "bytes_per_layer": {
+        "int4": 10473383366
+      },
+      "metadata": {
+        "architecture": "Mixture-of-Experts (MoE)",
+        "total_parameters": "1T",
+        "activated_parameters": "32B",
+        "num_layers": 61,
+        "context_length": 256000,
+        "native_quantization": "int4",
+        "download_size_gb": 595,
+        "recommended_short_name": "kimi-k2.7",
+        "recommended_engines": [
+          "vLLM",
+          "SGLang",
+          "KTransformers"
+        ]
+      }
+    }
+  }
+}
--- a/packages/tracker/meshnet_tracker/server.py
+++ b/packages/tracker/meshnet_tracker/server.py
@@ -31,6 +31,7 @@ import time
 import urllib.parse
 import urllib.request
 import uuid
+from importlib.resources import files
 from typing import Any

 from .gossip import NodeGossip
@@ -50,6 +51,23 @@ def derive_relay_url_from_public_tracker_url(url: str | None) -> str | None:
    return f"wss://{parsed.netloc}/ws"


+def _load_model_presets_from_data() -> dict[str, dict]:
+    """Load recommended model presets from package JSON data."""
+    try:
+        raw = files("meshnet_tracker").joinpath("model_presets.json").read_text()
+        data = json.loads(raw)
+    except Exception:
+        return {}
+    models = data.get("models", {})
+    if not isinstance(models, dict):
+        return {}
+    return {
+        str(name): preset
+        for name, preset in models.items()
+        if isinstance(preset, dict)
+    }
+
+
 DEFAULT_MODEL_PRESETS: dict[str, dict] = {
    "stub-model": {
        "layers_start": 0,
@@ -61,6 +79,7 @@ DEFAULT_MODEL_PRESETS: dict[str, dict] = {
        "layers_end": 11,
        "bytes_per_layer": {"bfloat16": 30 * 1024 * 1024, "int8": 15 * 1024 * 1024, "nf4": 8 * 1024 * 1024},
    },
+    **_load_model_presets_from_data(),
 }

 DEFAULT_VRAM_BYTES = 8 * 1024 * 1024 * 1024
@@ -83,6 +102,27 @@ def _model_aliases(model: str | None) -> set[str]:
    return aliases


+def _preset_aliases(name: str, preset: dict | None) -> set[str]:
+    aliases = _model_aliases(name)
+    if not preset:
+        return aliases
+    hf_repo = preset.get("hf_repo")
+    if isinstance(hf_repo, str):
+        aliases |= _model_aliases(hf_repo)
+    for alias in preset.get("aliases", []) or []:
+        if isinstance(alias, str):
+            aliases |= _model_aliases(alias)
+    return aliases
+
+
+def _resolve_model_preset(model_presets: dict, model: str) -> tuple[str, dict] | tuple[None, None]:
+    requested = _model_aliases(model)
+    for name, preset in model_presets.items():
+        if requested & _preset_aliases(name, preset):
+            return name, preset
+    return None, None
+
+
 def _node_matches_model(node: "_NodeEntry", model: str) -> bool:
    requested = _model_aliases(model)
    if not requested:
@@ -90,6 +130,11 @@ def _node_matches_model(node: "_NodeEntry", model: str) -> bool:
    return bool(requested & (_model_aliases(node.model) | _model_aliases(node.hf_repo)))


+def _node_matches_preset(node: "_NodeEntry", name: str, preset: dict) -> bool:
+    requested = _preset_aliases(name, preset)
+    return bool(requested & (_model_aliases(node.model) | _model_aliases(node.hf_repo)))
+
+
 class _RollingCounter:
    """Circular-bucket request counter.

@@ -258,7 +303,7 @@ class _StatsCollector:
 class _NodeEntry:
    __slots__ = (
        "node_id", "endpoint", "shard_start", "shard_end",
-        "model", "hf_repo", "num_layers", "shard_checksum", "hardware_profile", "wallet_address",
+        "model", "hf_repo", "num_layers", "model_metadata", "shard_checksum", "hardware_profile", "wallet_address",
        "score", "vram_bytes", "ram_bytes", "quantizations", "max_loaded_shards",
        "benchmark_tokens_per_sec", "quantization", "managed_assignment",
        "pending_directives", "last_heartbeat", "tracker_mode",
@@ -292,6 +337,7 @@ class _NodeEntry:
        tracker_mode: bool = False,
        hf_repo: str | None = None,
        num_layers: int | None = None,
+        model_metadata: dict | None = None,
        relay_addr: str | None = None,
        cert_fingerprint: str | None = None,
        peer_id: str | None = None,
@@ -315,6 +361,7 @@ class _NodeEntry:
        self.tracker_mode = tracker_mode
        self.hf_repo = hf_repo
        self.num_layers = num_layers
+        self.model_metadata = dict(model_metadata or {})
        self.relay_addr = relay_addr
        self.cert_fingerprint = cert_fingerprint
        self.peer_id = peer_id
@@ -467,6 +514,72 @@ def _node_capacity_summary(node: _NodeEntry, preset: dict | None = None) -> dict
    return summary


+def _node_memory_budget_for_preset(node: _NodeEntry, preset: dict | None = None) -> int:
+    budget, _source = _node_memory_budget_bytes(node)
+    if preset is None:
+        return int(budget * 0.8)
+    return _node_layer_capacity(node, preset) * max(1, next(iter(_preset_bytes_per_layer(preset).values())))
+
+
+def _pool_summary(nodes: list[_NodeEntry], preset: dict | None = None) -> dict:
+    total_vram = sum(max(0, node.vram_bytes) for node in nodes)
+    total_ram = sum(max(0, node.ram_bytes) for node in nodes)
+    total_budget = sum(_node_memory_budget_bytes(node)[0] for node in nodes)
+    effective_budget = sum(_node_memory_budget_for_preset(node, preset) for node in nodes)
+    return {
+        "node_count": len(nodes),
+        "total_vram_bytes": total_vram,
+        "total_ram_bytes": total_ram,
+        "total_memory_budget_bytes": total_budget,
+        "effective_assignable_memory_bytes": effective_budget,
+        "total_benchmark_tokens_per_sec": round(sum(node.benchmark_tokens_per_sec for node in nodes), 4),
+        "total_effective_throughput": round(sum(_effective_throughput(node) for node in nodes), 4),
+    }
+
+
+def _deployment_summary(nodes: list[_NodeEntry], preset: dict | None) -> dict:
+    if preset is None:
+        return {"recommended": False}
+    pool = _pool_summary(nodes, preset)
+    required = int(preset.get("required_model_bytes", 0) or 0)
+    deployable = required > 0 and pool["effective_assignable_memory_bytes"] >= required
+    missing = max(0, required - pool["effective_assignable_memory_bytes"]) if required > 0 else 0
+    return {
+        "recommended": bool(preset.get("recommended", False)),
+        "status": preset.get("deployment_status", "available"),
+        "required_model_bytes": required or None,
+        "download_size_bytes": preset.get("download_size_bytes"),
+        "native_quantization": preset.get("native_quantization"),
+        "pool": pool,
+        "deployable": deployable,
+        "missing_effective_memory_bytes": missing,
+    }
+
+
+def _max_layers_for_memory(memory_mb: int, total_layers: int, preset: dict | None = None) -> int:
+    if total_layers <= 0:
+        return 0
+    if memory_mb <= 0:
+        return max(1, total_layers // 2)
+    bytes_per_layer = next(iter(_preset_bytes_per_layer(preset).values())) if preset is not None else 30 * 1024 * 1024
+    return min(
+        total_layers,
+        max(1, int(((memory_mb * 1024 * 1024) * 0.8) // bytes_per_layer)),
+    )
+
+
+def _model_metadata_from_nodes(nodes: list[_NodeEntry]) -> dict:
+    metadata: dict = {}
+    for node in nodes:
+        if node.model_metadata:
+            metadata.update(node.model_metadata)
+    if "num_layers" not in metadata:
+        layers = [node.num_layers for node in nodes if node.num_layers is not None]
+        if layers:
+            metadata["num_layers"] = max(layers)
+    return metadata
+
+
 def _coverage_map(
    nodes: list[_NodeEntry],
    required_start: int,
@@ -608,10 +721,13 @@ def _nodes_and_bounds_for_model(
    server: "_TrackerHTTPServer",
    model: str,
 ) -> tuple[list[_NodeEntry], int, int] | None:
-    preset = server.model_presets.get(model)
+    resolved_name, preset = _resolve_model_preset(server.model_presets, model)
    if preset is not None:
        required_start, required_end = _preset_layer_bounds(preset)
-        return [node for node in server.registry.values() if node.model == model], required_start, required_end
+        return [
+            node for node in server.registry.values()
+            if _node_matches_preset(node, resolved_name, preset)  # type: ignore[arg-type]
+        ], required_start, required_end

    nodes = [
        node for node in server.registry.values()
@@ -668,12 +784,15 @@ def _purge_expired_nodes_locked(server: "_TrackerHTTPServer") -> list[str]:


 def _rebalance_model_locked(server: "_TrackerHTTPServer", model: str) -> None:
-    preset = server.model_presets.get(model)
+    resolved_name, preset = _resolve_model_preset(server.model_presets, model)
    if preset is None:
        return
    required_start, required_end = _preset_layer_bounds(preset)
    total_layers = required_end - required_start + 1
-    model_nodes = [node for node in server.registry.values() if node.model == model]
+    model_nodes = [
+        node for node in server.registry.values()
+        if _node_matches_preset(node, resolved_name, preset)  # type: ignore[arg-type]
+    ]
    managed_nodes = [node for node in model_nodes if node.managed_assignment]
    if not managed_nodes:
        return
@@ -1018,8 +1137,8 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
        data = []
        seen_ids: set[str] = set()
        for name, preset in server.model_presets.items():
-            model_nodes = [node for node in alive if node.model == name]
-            if not model_nodes:
+            model_nodes = [node for node in alive if _node_matches_preset(node, name, preset)]
+            if not model_nodes and not preset.get("recommended"):
                continue
            required_start, required_end = _preset_layer_bounds(preset)
            coverage = _coverage_percentage(
@@ -1031,6 +1150,9 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
            hf_repo = preset.get("hf_repo")
            if hf_repo and hf_repo not in aliases:
                aliases.append(hf_repo)
+            for alias in preset.get("aliases", []) or []:
+                if isinstance(alias, str) and alias not in aliases:
+                    aliases.append(alias)
            data.append({
                "id": name,
                "object": "model",
@@ -1039,9 +1161,14 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
                "name": name,
                "hf_repo": hf_repo,
                "aliases": aliases,
+                "metadata": dict(preset.get("metadata") or _model_metadata_from_nodes(model_nodes)),
+                "recommended": bool(preset.get("recommended", False)),
+                "deployment": _deployment_summary(alive, preset),
                "shard_coverage_percentage": coverage,
            })
            seen_ids.add(name)
+            if hf_repo:
+                seen_ids.add(hf_repo)

        hf_model_ids = sorted({
            node.hf_repo or node.model
@@ -1076,6 +1203,7 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
                "name": short_names[0] if short_names else model_id,
                "hf_repo": model_id if any(node.hf_repo == model_id for node in model_nodes) else None,
                "aliases": aliases,
+                "metadata": _model_metadata_from_nodes(model_nodes),
                "shard_coverage_percentage": _coverage_percentage(
                    model_nodes,
                    required_start,
@@ -1106,14 +1234,17 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
    def _handle_tracker_nodes(self, model: str):
        """Return nodes registered with tracker_mode=True whose shard starts at layer 0."""
        server: _TrackerHTTPServer = self.server  # type: ignore[assignment]
-        preset = server.model_presets.get(model)
+        resolved_name, preset = _resolve_model_preset(server.model_presets, model)
        if preset is None:
            self._send_json(404, {"error": f"unknown model preset: {model!r}"})
            return
        required_start, _ = _preset_layer_bounds(preset)
        with server.lock:
            self._purge_expired_nodes()
-            alive = [node for node in server.registry.values() if node.model == model]
+            alive = [
+                node for node in server.registry.values()
+                if _node_matches_preset(node, resolved_name, preset)  # type: ignore[arg-type]
+            ]
        if server.contracts is not None:
            alive = [
                node for node in alive
@@ -1126,7 +1257,7 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
            and node.tracker_mode
        ]
        self._send_json(200, {
-            "model": model,
+            "model": resolved_name,
            "tracker_nodes": [
                {
                    "node_id": node.node_id,
@@ -1155,6 +1286,18 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):

        self._send_json(200, {
            "relay_url": server.relay_url,
+            "pool": _pool_summary(nodes),
+            "recommended_models": [
+                {
+                    "id": name,
+                    "hf_repo": preset.get("hf_repo"),
+                    "aliases": list(preset.get("aliases", []) or []),
+                    "metadata": dict(preset.get("metadata") or {}),
+                    "deployment": _deployment_summary(nodes, preset),
+                }
+                for name, preset in server.model_presets.items()
+                if preset.get("recommended")
+            ],
            "nodes": [
                {
                    "node_id": node.node_id,
@@ -1163,6 +1306,8 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
                    "peer_id": node.peer_id,
                    "model": node.model,
                    "hf_repo": node.hf_repo,
+                    "num_layers": node.num_layers,
+                    "model_metadata": dict(node.model_metadata),
                    "shard_start": node.shard_start,
                    "shard_end": node.shard_end,
                    "tracker_mode": node.tracker_mode,
@@ -1520,6 +1665,12 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
            except (TypeError, ValueError):
                self._send_json(400, {"error": "num_layers must be an integer"})
                return
+        model_metadata = body.get("model_metadata", {})
+        if model_metadata is None:
+            model_metadata = {}
+        if not isinstance(model_metadata, dict):
+            self._send_json(400, {"error": "model_metadata must be an object"})
+            return
        relay_addr = body.get("relay_addr") or None
        cert_fingerprint = body.get("cert_fingerprint") or None
        peer_id = body.get("peer_id") or None
@@ -1552,6 +1703,7 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
            tracker_mode=tracker_mode,
            hf_repo=hf_repo,
            num_layers=num_layers,
+            model_metadata=model_metadata,
            relay_addr=relay_addr,
            cert_fingerprint=cert_fingerprint,
            peer_id=peer_id,
@@ -1743,7 +1895,7 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
        else:
            model = model_list[0]

-        preset = server.model_presets.get(model)
+        resolved_name, preset = _resolve_model_preset(server.model_presets, model)
        if preset is None:
            self._send_json(404, {"error": f"unknown model preset: {model!r}"})
            return
@@ -1752,7 +1904,10 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):

        with server.lock:
            self._purge_expired_nodes()
-            alive = [node for node in server.registry.values() if node.model == model]
+            alive = [
+                node for node in server.registry.values()
+                if _node_matches_preset(node, resolved_name, preset)  # type: ignore[arg-type]
+            ]
        if server.contracts is not None:
            alive = [
                node for node in alive
@@ -1805,7 +1960,7 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
        peers = [
            {"endpoint": node.endpoint, "checksum": node.shard_checksum}
            for node in alive
-            if node.model == model
+            if _node_matches_preset(node, resolved_name, preset)  # type: ignore[arg-type]
            and node.shard_start == shard_start
            and node.shard_end == shard_end
            and node.shard_checksum
@@ -1814,7 +1969,7 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
        self._send_json(200, {
            "shard_start": shard_start,
            "shard_end": shard_end,
-            "model": model,
+            "model": resolved_name,
            "model_layers_end": required_end,
            "peers": peers,
            **({"hf_repo": preset["hf_repo"]} if "hf_repo" in preset else {}),
@@ -1861,6 +2016,37 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
        ]

        if not hf_nodes:
+            resolved_name = None
+            preset = None
+            if filter_repo:
+                resolved_name, preset = _resolve_model_preset(server.model_presets, filter_repo)
+            else:
+                deployable = [
+                    (name, preset)
+                    for name, preset in server.model_presets.items()
+                    if preset.get("recommended") and _deployment_summary(all_nodes, preset)["deployable"]
+                ]
+                if deployable:
+                    resolved_name, preset = deployable[0]
+            if preset is not None and preset.get("hf_repo"):
+                required_start, required_end = _preset_layer_bounds(preset)
+                total_l = required_end - required_start + 1
+                memory_mb = vram_mb if vram_mb > 0 else ram_mb
+                max_layers = _max_layers_for_memory(memory_mb, total_l, preset)
+                shard_start = required_start
+                shard_end = min(required_end, shard_start + max_layers - 1)
+                self._send_json(200, {
+                    "hf_repo": preset["hf_repo"],
+                    "model": resolved_name,
+                    "shard_start": shard_start,
+                    "shard_end": shard_end,
+                    "num_layers": total_l,
+                    "gap_found": True,
+                    "price_per_token": 0.0,
+                    "deployment": _deployment_summary(all_nodes, preset),
+                })
+                return
+
            msg = (
                f"no HF-model nodes registered for {filter_repo!r}"
                if filter_repo
@@ -1929,15 +2115,13 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
        # Capacity: use the same 80%-of-memory rule as registered node planning.
        total_l = best_num_layers
        memory_mb = vram_mb if vram_mb > 0 else ram_mb
+        _resolved_name, best_preset = _resolve_model_preset(server.model_presets, str(best_repo))
        if memory_mb > 0:
-            max_layers = min(
-                total_l,
-                max(1, int(((memory_mb * 1024 * 1024) * 0.8) // (30 * 1024 * 1024))),
-            )
+            max_layers = _max_layers_for_memory(memory_mb, total_l, best_preset)
        elif device == "cuda" and vram_mb >= 8192:
            max_layers = total_l
        else:
-            max_layers = max(1, total_l // 2)
+            max_layers = _max_layers_for_memory(memory_mb, total_l, best_preset)

        shard_start = best_gap_start
        shard_end = min(total_l - 1, shard_start + max_layers - 1)
@@ -1960,13 +2144,16 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
            return

        model = model_list[0]
-        preset = server.model_presets.get(model)
+        resolved_name, preset = _resolve_model_preset(server.model_presets, model)

        with server.lock:
            self._purge_expired_nodes()
            if preset is not None:
                # Preset-based routing (stub-model system).
-                alive = [node for node in server.registry.values() if node.model == model]
+                alive = [
+                    node for node in server.registry.values()
+                    if _node_matches_preset(node, resolved_name, preset)  # type: ignore[arg-type]
+                ]
                required_start, required_end = _preset_layer_bounds(preset)
            else:
                # HF model routing: match by hf_repo (full) or model short name.
@@ -2013,6 +2200,9 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
                    "shard_start": e.shard_start,
                    "shard_end": e.shard_end,
                    "model": e.model,
+                    "hf_repo": e.hf_repo,
+                    "num_layers": e.num_layers,
+                    "model_metadata": dict(e.model_metadata),
                    "shard_checksum": e.shard_checksum,
                    "score": e.score,
                }
@@ -2038,7 +2228,7 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
            return

        model = model_list[0]
-        preset = server.model_presets.get(model)
+        resolved_name, preset = _resolve_model_preset(server.model_presets, model)
        if preset is None:
            self._send_json(404, {"error": f"unknown model preset: {model!r}"})
            return
@@ -2047,7 +2237,10 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):

        with server.lock:
            self._purge_expired_nodes()
-            candidates = [node for node in server.registry.values() if node.model == model]
+            candidates = [
+                node for node in server.registry.values()
+                if _node_matches_preset(node, resolved_name, preset)  # type: ignore[arg-type]
+            ]
        if server.contracts is not None:
            candidates = [
                node for node in candidates
@@ -2208,6 +2401,7 @@ class TrackerServer:
            tracker_mode=bool(payload.get("tracker_mode", False)),
            hf_repo=payload.get("hf_repo"),
            num_layers=int(payload["num_layers"]) if payload.get("num_layers") is not None else None,
+            model_metadata=payload.get("model_metadata") if isinstance(payload.get("model_metadata"), dict) else None,
        )
        with self._lock:
            self._registry[node_id] = entry
--- a/packages/tracker/pyproject.toml
+++ b/packages/tracker/pyproject.toml
@@ -18,3 +18,6 @@ meshnet-tracker = "meshnet_tracker.cli:main"
 [tool.setuptools.packages.find]
 where = ["."]
 include = ["meshnet_tracker*"]
+
+[tool.setuptools.package-data]
+meshnet_tracker = ["*.json"]
--- a/tests/test_node_startup.py
+++ b/tests/test_node_startup.py
@@ -197,6 +197,65 @@ def test_benchmark_throughput_is_registered_in_payload(monkeypatch, tmp_path):
    assert captured["hardware_profile"]["benchmark_ok"] is True


+def test_real_model_startup_passes_download_dir_and_kimi_metadata(monkeypatch, tmp_path):
+    import meshnet_node.startup as startup_mod
+
+    captured_registration: dict = {}
+    captured_torch_kwargs: dict = {}
+
+    class FakeBackend:
+        total_layers = 61
+
+    class FakeNode:
+        backend = FakeBackend()
+
+        def __init__(self, **kwargs):
+            captured_torch_kwargs.update(kwargs)
+
+        def start(self):
+            return 7099
+
+        def stop(self):
+            pass
+
+        def apply_tracker_directives(self, directives):
+            return None
+
+    monkeypatch.setattr(
+        startup_mod,
+        "detect_hardware",
+        lambda: {"device": "cpu", "gpu_name": None, "vram_mb": 0, "ram_mb": 16384},
+    )
+    monkeypatch.setattr(startup_mod, "benchmark_throughput_checked", lambda _device: (42.5, True, None))
+    monkeypatch.setattr(startup_mod, "TorchNodeServer", FakeNode)
+    monkeypatch.setattr(startup_mod, "load_or_create_wallet", lambda **_kw: (b"", b"", "wallet-kimi"))
+    monkeypatch.setattr(startup_mod, "_get_json", lambda _url, timeout=10.0: {"relay_url": None, "nodes": []})
+    monkeypatch.setattr(
+        startup_mod,
+        "_post_json",
+        lambda _url, payload, timeout=10.0: (
+            captured_registration.update(payload) or {"node_id": "node-kimi"}
+        ),
+    )
+    monkeypatch.setattr(startup_mod, "_start_heartbeat", lambda *a, **kw: None)
+
+    cache_dir = tmp_path / "models"
+    node = run_startup(
+        tracker_url="http://localhost:8080",
+        model_id="unsloth/Kimi-K2.7-Code",
+        shard_start=0,
+        shard_end=60,
+        wallet_path=tmp_path / "wallet.json",
+        cache_dir=cache_dir,
+    )
+    node.stop()
+
+    assert captured_torch_kwargs["cache_dir"] == cache_dir
+    assert captured_registration["model_metadata"]["total_parameters"] == "1T"
+    assert captured_registration["model_metadata"]["activated_parameters"] == "32B"
+    assert captured_registration["model_metadata"]["context_length"] == 256000
+
+
 def test_cuda_benchmark_failure_is_registered_for_inventory_only_gpu(monkeypatch, tmp_path, capsys):
    import meshnet_node.startup as startup_mod

--- a/tests/test_tracker_routing.py
+++ b/tests/test_tracker_routing.py
@@ -50,6 +50,151 @@ def test_tracker_send_json_ignores_broken_pipe_after_client_disconnect():
    _TrackerHandler._send_json(DummyHandler(), 200, {"ok": True})


+def test_tracker_exposes_registered_model_metadata():
+    tracker = TrackerServer()
+    port = tracker.start()
+    url = f"http://127.0.0.1:{port}"
+    try:
+        _post_json(
+            f"{url}/v1/nodes/register",
+            {
+                "endpoint": "http://127.0.0.1:7100",
+                "model": "Kimi-K2.7-Code",
+                "hf_repo": "unsloth/Kimi-K2.7-Code",
+                "num_layers": 61,
+                "shard_start": 0,
+                "shard_end": 60,
+                "hardware_profile": {},
+                "model_metadata": {
+                    "total_parameters": "1T",
+                    "activated_parameters": "32B",
+                    "context_length": 256000,
+                },
+            },
+        )
+
+        models = _get_json(f"{url}/v1/models")
+        network_map = _get_json(f"{url}/v1/network/map")
+    finally:
+        tracker.stop()
+
+    kimi = next(model for model in models["data"] if model["id"] == "kimi-k2.7")
+    assert kimi["hf_repo"] == "unsloth/Kimi-K2.7-Code"
+    assert kimi["metadata"]["total_parameters"] == "1T"
+    assert kimi["metadata"]["activated_parameters"] == "32B"
+    assert kimi["metadata"]["num_layers"] == 61
+    registered = network_map["nodes"][0]
+    assert registered["num_layers"] == 61
+    assert registered["model_metadata"]["context_length"] == 256000
+
+
+def test_tracker_lists_recommended_kimi_before_nodes_register():
+    tracker = TrackerServer()
+    port = tracker.start()
+    url = f"http://127.0.0.1:{port}"
+    try:
+        models = _get_json(f"{url}/v1/models")
+        network_map = _get_json(f"{url}/v1/network/map")
+    finally:
+        tracker.stop()
+
+    kimi = next(model for model in models["data"] if model["id"] == "kimi-k2.7")
+    assert kimi["hf_repo"] == "unsloth/Kimi-K2.7-Code"
+    assert "Kimi-K2.7-Code" in kimi["aliases"]
+    assert kimi["metadata"]["download_size_gb"] == 595
+    assert kimi["deployment"]["native_quantization"] == "int4"
+    assert kimi["deployment"]["deployable"] is False
+    assert network_map["pool"]["node_count"] == 0
+    assert network_map["recommended_models"][0]["id"] == "kimi-k2.7"
+
+
+def test_network_map_exposes_pool_size_and_speed_summary():
+    tracker = TrackerServer()
+    port = tracker.start()
+    url = f"http://127.0.0.1:{port}"
+    try:
+        _post_json(
+            f"{url}/v1/nodes/register",
+            {
+                "endpoint": "http://127.0.0.1:7101",
+                "model": "inventory-a",
+                "vram_bytes": 10_000,
+                "ram_bytes": 20_000,
+                "benchmark_tokens_per_sec": 4.0,
+                "hardware_profile": {},
+            },
+        )
+        _post_json(
+            f"{url}/v1/nodes/register",
+            {
+                "endpoint": "http://127.0.0.1:7102",
+                "model": "inventory-b",
+                "vram_bytes": 0,
+                "ram_bytes": 30_000,
+                "benchmark_tokens_per_sec": 6.0,
+                "hardware_profile": {},
+            },
+        )
+
+        network_map = _get_json(f"{url}/v1/network/map")
+    finally:
+        tracker.stop()
+
+    pool = network_map["pool"]
+    assert pool["node_count"] == 2
+    assert pool["total_vram_bytes"] == 10_000
+    assert pool["total_ram_bytes"] == 50_000
+    assert pool["total_memory_budget_bytes"] == 40_000
+    assert pool["total_benchmark_tokens_per_sec"] == 10.0
+    assert pool["total_effective_throughput"] == 10.0
+
+
+def test_recommended_kimi_becomes_deployable_when_pool_is_large_enough():
+    tracker = TrackerServer()
+    port = tracker.start()
+    url = f"http://127.0.0.1:{port}"
+    gib = 1024 * 1024 * 1024
+    try:
+        for idx in range(2):
+            _post_json(
+                f"{url}/v1/nodes/register",
+                {
+                    "endpoint": f"http://127.0.0.1:{7200 + idx}",
+                    "model": f"inventory-{idx}",
+                    "vram_bytes": 0,
+                    "ram_bytes": 400 * gib,
+                    "benchmark_tokens_per_sec": 5.0,
+                    "hardware_profile": {},
+                },
+            )
+
+        models = _get_json(f"{url}/v1/models")
+    finally:
+        tracker.stop()
+
+    kimi = next(model for model in models["data"] if model["id"] == "kimi-k2.7")
+    assert kimi["deployment"]["deployable"] is True
+    assert kimi["deployment"]["missing_effective_memory_bytes"] == 0
+
+
+def test_network_assign_can_start_recommended_kimi_by_short_name():
+    tracker = TrackerServer()
+    port = tracker.start()
+    url = f"http://127.0.0.1:{port}"
+    try:
+        assignment = _get_json(
+            f"{url}/v1/network/assign?device=cpu&ram_mb=204800&hf_repo=kimi-k2.7"
+        )
+    finally:
+        tracker.stop()
+
+    assert assignment["model"] == "kimi-k2.7"
+    assert assignment["hf_repo"] == "unsloth/Kimi-K2.7-Code"
+    assert assignment["num_layers"] == 61
+    assert assignment["shard_start"] == 0
+    assert 0 <= assignment["shard_end"] < 60
+
+
 def test_tracker_serves_health_while_proxy_request_is_in_flight():
    """Long inference proxy requests must not block heartbeats/health checks."""