`phimoe.py` subscripts `rope_scaling` unconditionally, so `Phi-mini-MoE-instruct` dies with a bare `TypeError`

[hijohnnylin]

• AI Usage Notice: We found this with Opus. Code is easily reproducible below.

---

microsoft/Phi-mini-MoE-instruct has "rope_scaling": null in its config.json. phimoe.py
reads four keys out of that field while building attention, so loading the checkpoint raises before
any weight is touched.

Versions: mlx-lm 0.31.3, mlx 0.32.2, Apple M3 Max.

Runnable repro

Part A needs no download and no checkpoint at all: ModelArgs.rope_scaling already defaults to
None, and the only code that reads it subscripts it four times, so the port contradicts its own
declared default. Part B is the checkpoint that made this visible.

pip install "mlx-lm>=0.31.3"
python repro_phimoe.py          # Part A only -- instant, no download
python repro_phimoe.py --real   # adds Part B (~7.6 GB)

"""Repro: mlx-lm's phimoe port cannot build a model whose `rope_scaling` is null.

Part A needs no download and no checkpoint -- `ModelArgs.rope_scaling` already defaults to
None, and the code that reads it cannot accept None, so the port contradicts its own default.
Part B is the real checkpoint (~7.6 GB) that made this visible.

    pip install "mlx-lm>=0.31.3"
    python repro_phimoe.py          # Part A only
    python repro_phimoe.py --real   # adds Part B
"""

import sys
import traceback

from mlx_lm.models import phimoe

print("=" * 78)
print("A. mlx-lm's own default, no checkpoint and no download")
print("=" * 78)

args = phimoe.ModelArgs(
    # Small so the failure is quick; the shape has nothing to do with the bug.
    hidden_size=64,
    intermediate_size=64,
    num_hidden_layers=1,
    num_attention_heads=4,
    num_key_value_heads=4,
    vocab_size=32,
    num_local_experts=2,
    num_experts_per_tok=1,
)
print(f"ModelArgs.rope_scaling = {args.rope_scaling!r}   <- the dataclass's declared default")

try:
    phimoe.Model(args)
except TypeError:
    frames = traceback.extract_tb(sys.exc_info()[2])
    last = frames[-1]
    print(f"\n{type(sys.exc_info()[1]).__name__}: {sys.exc_info()[1]}")
    print(f"  raised at {last.filename.split('/')[-1]}:{last.lineno}")
    print(f"  line:     {last.line}")
    print("\nThe field defaults to None and the only code that reads it subscripts it four times.")
else:
    raise SystemExit("no error -- this mlx-lm version has been fixed")

if "--real" not in sys.argv:
    print("\n(pass --real to also run Part B against microsoft/Phi-mini-MoE-instruct)")
    raise SystemExit(0)

print()
print("=" * 78)
print("B. the checkpoint this was found on: microsoft/Phi-mini-MoE-instruct")
print("=" * 78)
print('Its config.json carries "rope_scaling": null, so the same line is reached.')

from mlx_lm import load  # noqa: E402

try:
    load("microsoft/Phi-mini-MoE-instruct")
except TypeError:
    frames = traceback.extract_tb(sys.exc_info()[2])
    ours = [f for f in frames if "mlx_lm" in (f.filename or "")]
    print(f"\n{type(sys.exc_info()[1]).__name__}: {sys.exc_info()[1]}")
    print(f"  raised at {ours[-1].filename.split('/')[-1]}:{ours[-1].lineno}")
    print(f"\nmlx_lm frames in the traceback: {len(ours)} of {len(frames)}")
    print("Nothing in the message names `rope_scaling`, the config, or the checkpoint, which is")
    print("why this reads as a bug in whatever called load().")
else:
    raise SystemExit("no error -- this mlx-lm version has been fixed")

Output on mlx-lm 0.31.3 / mlx 0.32.2, Apple M3 Max:

==============================================================================
A. mlx-lm's own default, no checkpoint and no download
==============================================================================
ModelArgs.rope_scaling = None   <- the dataclass's declared default

TypeError: 'NoneType' object is not subscriptable
  raised at phimoe.py:53
  line:     short_factor=args.rope_scaling["short_factor"],

The field defaults to None and the only code that reads it subscripts it four times.

==============================================================================
B. the checkpoint this was found on: microsoft/Phi-mini-MoE-instruct
==============================================================================
Its config.json carries "rope_scaling": null, so the same line is reached.

TypeError: 'NoneType' object is not subscriptable
  raised at phimoe.py:53

mlx_lm frames in the traceback: 6 of 7

Cause

mlx_lm/models/phimoe.py, in Attention.__init__:

self.rope = SuScaledRoPE(
    head_dim,
    base=args.rope_theta,
    max_position_embeddings=args.max_position_embeddings,
    original_max_position_embeddings=args.original_max_position_embeddings,
    short_factor=args.rope_scaling["short_factor"],   # <- line 53
    long_factor=args.rope_scaling["long_factor"],
    short_mscale=args.rope_scaling["short_mscale"],
    long_mscale=args.rope_scaling["long_mscale"],
)

rope_scaling is optional in the Phi-MoE config, and this is the only place it is read. A config
that omits it or sets it to null means plain RoPE.

The traceback is the other half of the problem: it contains only mlx-lm frames and a bare
TypeError with no field name, no config key and no checkpoint name, so it reads as a bug in
whatever called load(). It cost a release's worth of misattribution here before I read the port.

Suggested fix

Fall back to plain RoPE when the field is absent:

scaling = args.rope_scaling
if scaling is None:
    self.rope = nn.RoPE(head_dim, traditional=False, base=args.rope_theta)
else:
    self.rope = SuScaledRoPE(...)

A clearer failure would help even without the fallback — raising ValueError naming
rope_scaling and the checkpoint would have made this a one-minute diagnosis rather than an
afternoon.

I am happy to open a PR if that is useful.