fix qwen moe tensor sharding (#1604)

## Motivation

<!-- Why is this change needed? What problem does it solve? -->
<!-- If it fixes an open issue, please link to the issue here -->

## Changes

<!-- Describe what you changed in detail -->

## Why It Works

<!-- Explain why your approach solves the problem -->

## Test Plan

### Manual Testing
<!-- Hardware: (e.g., MacBook Pro M1 Max 32GB, Mac Mini M2 16GB,
connected via Thunderbolt 4) -->
<!-- What you did: -->
<!-- - -->

### Automated Testing
<!-- Describe changes to automated tests, or how existing tests cover
this change -->
<!-- - -->

src/exo/main.py

@@ -261,13 +261,6 @@ def main():
     if args.offline:
         logger.info("Running in OFFLINE mode — no internet checks, local models only")
 
-    # Set trust_remote_code override env var for runner subprocesses
-    if args.trust_remote_code:
-        os.environ["EXO_TRUST_REMOTE_CODE"] = "1"
-        logger.warning(
-            "--trust-remote-code enabled: models may execute arbitrary code during loading"
-        )
-
     # Set FAST_SYNCH override env var for runner subprocesses
     if args.fast_synch is True:
         os.environ["EXO_FAST_SYNCH"] = "on"
@@ -292,7 +285,6 @@ class Args(CamelCaseModel):
     no_downloads: bool = False
     offline: bool = False
     fast_synch: bool | None = None  # None = auto, True = force on, False = force off
-    trust_remote_code: bool = False
 
     @classmethod
     def parse(cls) -> Self:
@@ -344,11 +336,6 @@ class Args(CamelCaseModel):
             action="store_true",
             help="Run in offline/air-gapped mode: skip internet checks, use only pre-staged local models",
         )
-        parser.add_argument(
-            "--trust-remote-code",
-            action="store_true",
-            help="Allow models to execute custom code during tokenizer loading (security-sensitive, CLI-only)",
-        )
         fast_synch_group = parser.add_mutually_exclusive_group()
         fast_synch_group.add_argument(
             "--fast-synch",

src/exo/worker/engines/mlx/auto_parallel.py

@@ -852,6 +852,8 @@ class QwenShardingStrategy(TensorParallelShardingStrategy):
                 layer.self_attn.o_proj = self.sharded_to_all_linear(
                     layer.self_attn.o_proj
                 )
+                layer.self_attn.n_heads //= self.N
+                layer.self_attn.n_kv_heads //= self.N
             else:
                 assert isinstance(layer, Qwen3NextDecoderLayer)
                 if hasattr(layer, "linear_attn"):

src/exo/worker/engines/mlx/constants.py

@@ -13,6 +13,5 @@ KV_CACHE_BITS: int | None = None
 
 DEFAULT_TOP_LOGPROBS: int = 5
 
-# True for built-in models with known model cards; custom models added via API default to False
-# and can be overridden with the --trust-remote-code CLI flag.
+# TODO: We should really make this opt-in, but Kimi requires trust_remote_code=True
 TRUST_REMOTE_CODE: bool = True

src/exo/worker/engines/mlx/utils_mlx.py

@@ -291,14 +291,10 @@ def shard_and_load(
 
 def get_tokenizer(model_path: Path, shard_metadata: ShardMetadata) -> TokenizerWrapper:
     """Load tokenizer for a model shard. Delegates to load_tokenizer_for_model_id."""
-    trust_remote_code = (
-        shard_metadata.model_card.trust_remote_code
-        or os.environ.get("EXO_TRUST_REMOTE_CODE") == "1"
-    )
     return load_tokenizer_for_model_id(
         shard_metadata.model_card.model_id,
         model_path,
-        trust_remote_code=trust_remote_code,
+        trust_remote_code=shard_metadata.model_card.trust_remote_code,
     )