personaplex: default server to 127.0.0.1; fix branch lint warnings

- server.py/README: default --host to 127.0.0.1 (localhost) instead of
  0.0.0.0, matching the documented SSH port-forward workflow and avoiding
  binding to all interfaces by default.
- Fix ruff warnings across the Moshi branch: D205 docstring summary/blank
  line (mimi.py), RUF005 list concatenation -> unpacking (golden script +
  integration test). Import sort/format auto-applied by lintrunner.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: Justin Chu <11205048+justinchuby@users.noreply.github.com>

Author: justinchuby

examples/personaplex/README.md

@@ -54,7 +54,7 @@ The server only loads ONNX models, so it can run in a lightweight
 pip install aiohttp onnxruntime-gpu numpy sentencepiece huggingface_hub
 python examples/personaplex/server.py \
     --model-dir output/personaplex/onnx --device cuda \
-    --host 0.0.0.0 --port 7681
+    --host 127.0.0.1 --port 7681
 ```
 
 On Ampere+/H200 GPUs ORT defaults to TF32 for fp32 matmuls, which can flip

examples/personaplex/server.py

@@ -30,7 +30,7 @@
     pip install aiohttp sentencepiece huggingface_hub
     python examples/personaplex/server.py \
         --model-dir output/personaplex/onnx --device cuda \
-        --host 0.0.0.0 --port 7681
+        --host 127.0.0.1 --port 7681
 
 Open ``http://localhost:7681`` (or port-forward the remote port over SSH:
 ``ssh -L 7681:localhost:7681 <host>``), set a persona / optional voice sample,
@@ -254,7 +254,7 @@ def main() -> None:
         action="store_true",
         help="keep CUDA TF32 (faster, lower precision) for fp32 matmuls",
     )
-    parser.add_argument("--host", default="0.0.0.0")
+    parser.add_argument("--host", default="127.0.0.1")
     parser.add_argument("--port", type=int, default=7681)
     parser.add_argument(
         "--tokenizer",

scripts/generate_moshi_lm_golden.py

@@ -57,7 +57,7 @@ def make_prev_tokens() -> list[int]:
     rng = np.random.RandomState(7)
     text_prev = int(rng.randint(0, TEXT_CARD))
     audio_prev = rng.randint(0, AUDIO_CARD, size=DEP_Q - 1).tolist()
-    return [text_prev] + audio_prev
+    return [text_prev, *audio_prev]
 
 
 def _resolve_lm(model: str) -> str:

src/mobius/models/mimi.py

@@ -35,7 +35,6 @@
 
 from __future__ import annotations
 
-
 import numpy as np
 import onnx_ir as ir
 import torch
@@ -236,8 +235,10 @@ def forward(self, op: OpBuilder, x: ir.Value):
 
 
 class _ELU(nn.Module):
-    """ELU activation (alpha=1.0); paramless, kept in the module list to
-    preserve Kyutai ``model.<i>`` indexing."""
+    """ELU activation (alpha=1.0).
+
+    Paramless; kept in the module list to preserve Kyutai ``model.<i>`` indexing.
+    """
 
     def forward(self, op: OpBuilder, x: ir.Value):
         return op.Elu(x, alpha=1.0)
@@ -687,8 +688,11 @@ def _interleaved_to_halfsplit(w: torch.Tensor, head_dim: int) -> torch.Tensor:
 def _convert_transformer_layer(
     out: dict[str, torch.Tensor], prefix: str, lprefix: str, sd: dict
 ) -> None:
-    """Convert one Kyutai transformer layer (``lprefix``) into mobius names
-    under ``prefix`` (the fully-qualified ``<enc|dec>oder.<...>.layers.<i>``)."""
+    """Convert one Kyutai transformer layer into mobius parameter names.
+
+    ``lprefix`` is the Kyutai source prefix; ``prefix`` is the fully-qualified
+    destination (``<enc|dec>oder.<...>.layers.<i>``).
+    """
     # Norms
     out[f"{prefix}.input_layernorm.weight"] = sd[f"{lprefix}.norm1.weight"]
     out[f"{prefix}.input_layernorm.bias"] = sd[f"{lprefix}.norm1.bias"]
@@ -711,8 +715,11 @@ def _convert_transformer_layer(
 
 
 def _codebook_embedding(sd: dict, prefix: str) -> torch.Tensor:
-    """Reconstruct a codebook embedding table from ``embedding_sum`` and
-    ``cluster_usage`` (Kyutai stores running statistics, not the table)."""
+    """Reconstruct a codebook embedding table from running statistics.
+
+    Kyutai stores ``embedding_sum`` and ``cluster_usage`` (running statistics),
+    not the embedding table itself.
+    """
     embedding_sum = sd[f"{prefix}.embedding_sum"].float()  # (bins, dim)
     cluster_usage = sd[f"{prefix}.cluster_usage"].float()  # (bins,)
     denom = cluster_usage.clamp(min=1e-8).unsqueeze(-1)

tests/moshi_lm_integration_test.py

@@ -64,7 +64,7 @@ def _make_prev_tokens() -> list[int]:
     rng = np.random.RandomState(7)
     text_prev = int(rng.randint(0, _TEXT_CARD))
     audio_prev = rng.randint(0, _AUDIO_CARD, size=_DEP_Q - 1).tolist()
-    return [text_prev] + audio_prev
+    return [text_prev, *audio_prev]
 
 
 def _find_onnx(root: str) -> str: