foo

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

@@ -4,7 +4,7 @@ from abc import ABC, abstractmethod
 from collections.abc import Callable
 from functools import partial
 from inspect import signature
-from typing import TYPE_CHECKING, Any, Protocol, cast
+from typing import TYPE_CHECKING, Any, cast
 
 import mlx.core as mx
 import mlx.nn as nn
@@ -13,11 +13,17 @@ from mlx.nn.layers.distributed import (
     shard_linear,
     sum_gradients,
 )
+from mlx_lm.models.deepseek_v3 import DeepseekV3MLP
 from mlx_lm.models.deepseek_v3 import Model as DeepseekV3Model
+from mlx_lm.models.deepseek_v32 import DeepseekV32MLP
 from mlx_lm.models.deepseek_v32 import Model as DeepseekV32Model
 from mlx_lm.models.glm4_moe import Model as Glm4MoeModel
 from mlx_lm.models.glm4_moe import MoE
 from mlx_lm.models.gpt_oss import GptOssMoeModel
+from mlx_lm.models.gpt_oss import Model as GptOssModel
+from mlx_lm.models.llama import Model as LlamaModel
+from mlx_lm.models.minimax import Model as MiniMaxModel
+from mlx_lm.models.ministral3 import Model as Ministral3Model
 from mlx_lm.models.qwen3_moe import Model as Qwen3MoeModel
 from mlx_lm.models.qwen3_moe import Qwen3MoeSparseMoeBlock
 from mlx_lm.models.qwen3_next import Model as Qwen3NextModel
@@ -60,28 +66,16 @@ def eval_with_timeout(
     finally:
         completed.set()
 
-
-class _LayerCallable(Protocol):
-    """Structural type that any compatible layer must satisfy.
-
-    We require a single positional input of type ``mx.array`` and an
-    ``mx.array`` output, while permitting arbitrary *args / **kwargs so this
-    protocol matches the vast majority of `mlx.nn.Module` subclasses.
-    """
-
-    def __call__(self, x: mx.array, *args: object, **kwargs: object) -> mx.array: ...
-
-
 class CustomMlxLayer(nn.Module):
     """Base class for replacing an MLX layer with a custom implementation."""
 
-    def __init__(self, original_layer: _LayerCallable):
+    def __init__(self, original_layer: nn.Module):
         super().__init__()
         object.__setattr__(self, "_original_layer", original_layer)
 
     @property
-    def original_layer(self) -> _LayerCallable:
-        return cast(_LayerCallable, object.__getattribute__(self, "_original_layer"))
+    def original_layer(self) -> nn.Module:
+        return cast(nn.Module, object.__getattribute__(self, "_original_layer"))
 
     # Calls __getattr__ for any attributes not found on nn.Module (e.g. use_sliding)
     if not TYPE_CHECKING:
@@ -94,53 +88,49 @@ class CustomMlxLayer(nn.Module):
                 return getattr(original_layer, name)
 
 
-class PipelineFirstLayer(CustomMlxLayer):
-    def __init__(
-        self,
-        original_layer: _LayerCallable,
-        r: int,
-        group: mx.distributed.Group,
-    ):
-        super().__init__(original_layer)
-        self.r: int = r
-        self.group = group
-
-    def __call__(self, x: mx.array, *args: object, **kwargs: object) -> mx.array:
-        if self.r != 0:
-            x = mx.distributed.recv_like(x, (self.r - 1), group=self.group)
-        return self.original_layer(x, *args, **kwargs)
 
 
-class PipelineLastLayer(CustomMlxLayer):
-    def __init__(
-        self,
-        original_layer: _LayerCallable,
-        r: int,
-        s: int,
-        group: mx.distributed.Group,
-    ):
-        super().__init__(original_layer)
-        self.r: int = r
-        self.s: int = s
-        self.group = group
-        self.original_layer_signature = signature(self.original_layer.__call__)
+def patch_pipeline_first_layer(pipeline_layer: nn.Module, group: mx.distributed.Group) -> nn.Module:
+    orig_call = cast(Callable[..., mx.array], type(pipeline_layer).__call__)
 
-    def __call__(self, x: mx.array, *args: object, **kwargs: object) -> mx.array:
-        cache = self.original_layer_signature.bind_partial(
-            x, *args, **kwargs
-        ).arguments.get("cache", None)
+    rank = group.rank()
+    class PatchedFirstLayer(nn.Module):
+        def __call__(self, x: mx.array, *args: object, **kwargs: object) -> mx.array:
+            if rank != 0:
+                x = mx.distributed.recv_like(x, (rank - 1), group=group)
+            return orig_call(x, *args, **kwargs)
 
-        output: mx.array = self.original_layer(x, *args, **kwargs)
+    pipeline_layer.__class__ = PatchedFirstLayer
+            
+    return pipeline_layer
 
-        if self.r != self.s - 1:
-            output = mx.distributed.send(
-                output, (self.r + 1) % self.s, group=self.group
-            )
-            if cache is not None:
-                cache.keys = mx.depends(cache.keys, output)  # type: ignore[reportUnknownMemberType]
+def patch_pipeline_last_layer(pipeline_layer: nn.Module, group: mx.distributed.Group) -> nn.Module:
+    orig_call = cast(Callable[..., mx.array], type(pipeline_layer).__call__)
+    orig_call_sig = signature(orig_call)
+    
 
-        return output
+    rank = group.rank()
+    size = group.size()
+    class PatchedLastLayer(nn.Module):
+        def __call__(self, x: mx.array, *args: object, **kwargs: object) -> mx.array:
+            cache = orig_call_sig.bind_partial(
+                x, *args, **kwargs
+            ).arguments.get("cache", None)
 
+            output: mx.array = orig_call(x, *args, **kwargs)
+
+            if rank != size - 1:
+                output = mx.distributed.send(
+                    output, (rank + 1) % size, group=group
+                )
+                if cache is not None:
+                    cache.keys = mx.depends(cache.keys, output)  # type: ignore[reportUnknownMemberType]
+
+            return output
+
+    pipeline_layer.__class__ = PatchedLastLayer
+            
+    return pipeline_layer
 
 def _inner_model(model: nn.Module) -> nn.Module:
     inner = getattr(model, "model", None)
@@ -154,13 +144,13 @@ def _inner_model(model: nn.Module) -> nn.Module:
     raise ValueError("Model must either have a 'model' or 'transformer' attribute")
 
 
-def _get_layers(inner_model_instance: nn.Module) -> list[_LayerCallable]:
+def _get_layers(inner_model_instance: nn.Module) -> list[nn.Module]:
     # Handle both model.layers and model.h cases
-    layers: list[_LayerCallable]
+    layers: list[nn.Module]
     if hasattr(inner_model_instance, "layers"):
-        layers = cast(list[_LayerCallable], inner_model_instance.layers)
+        layers = cast(list[nn.Module], inner_model_instance.layers)
     elif hasattr(inner_model_instance, "h"):
-        layers = cast(list[_LayerCallable], inner_model_instance.h)
+        layers = cast(list[nn.Module], inner_model_instance.h)
     else:
         raise ValueError("Model must have either a 'layers' or 'h' attribute")
 
@@ -185,15 +175,12 @@ def pipeline_auto_parallel(
     layers = _get_layers(inner_model_instance)
 
     start_layer, end_layer = model_shard_meta.start_layer, model_shard_meta.end_layer
-    device_rank, world_size = model_shard_meta.device_rank, model_shard_meta.world_size
 
     layers = layers[start_layer:end_layer]
-    layers[0] = PipelineFirstLayer(layers[0], device_rank, group=group)
-    layers[-1] = PipelineLastLayer(
+    layers[0] = patch_pipeline_first_layer(layers[0], group)
+    layers[-1] = patch_pipeline_last_layer(
         layers[-1],
-        device_rank,
-        world_size,
-        group=group,
+        group,
     )
 
     if isinstance(inner_model_instance, GptOssMoeModel):
@@ -335,7 +322,33 @@ def tensor_auto_parallel(
         except (AttributeError, TypeError, NameError):
             pass
 
-    if isinstance(model, (Qwen3MoeModel, Glm4MoeModel, Qwen3NextModel)):
+    if isinstance(model, (LlamaModel, Ministral3Model)):
+        logger.warning("shouldn't be hit - upstream sharding exists")
+        tensor_parallel_sharding_strategy = LlamaShardingStrategy(
+            group,
+            all_to_sharded_linear,
+            sharded_to_all_linear,
+            all_to_sharded_linear_in_place,
+            sharded_to_all_linear_in_place,
+        )
+    elif isinstance(model, (DeepseekV3Model, DeepseekV32Model)):
+        logger.warning("shouldn't be hit - upstream sharding exists")
+        tensor_parallel_sharding_strategy = DeepSeekShardingStrategy(
+            group,
+            all_to_sharded_linear,
+            sharded_to_all_linear,
+            all_to_sharded_linear_in_place,
+            sharded_to_all_linear_in_place,
+        )
+    elif isinstance(model, MiniMaxModel):
+        tensor_parallel_sharding_strategy = MiniMaxShardingStrategy(
+            group,
+            all_to_sharded_linear,
+            sharded_to_all_linear,
+            all_to_sharded_linear_in_place,
+            sharded_to_all_linear_in_place,
+        )
+    elif isinstance(model, (Qwen3MoeModel, Glm4MoeModel, Qwen3NextModel)):
         tensor_parallel_sharding_strategy = QwenShardingStrategy(
             group,
             all_to_sharded_linear,
@@ -343,6 +356,15 @@ def tensor_auto_parallel(
             all_to_sharded_linear_in_place,
             sharded_to_all_linear_in_place,
         )
+    elif isinstance(model, GptOssModel):
+        tensor_parallel_sharding_strategy = GptOssShardingStrategy(
+            group,
+            all_to_sharded_linear,
+            sharded_to_all_linear,
+            all_to_sharded_linear_in_place,
+            sharded_to_all_linear_in_place,
+        )
+
     else:
         raise ValueError(f"Unsupported model type: {type(model)}")
 
@@ -377,7 +399,35 @@ class TensorParallelShardingStrategy(ABC):
     ) -> nn.Module: ...
 
 
-def _set_layers(model: nn.Module, layers: list[_LayerCallable]) -> None:
+class LlamaShardingStrategy(TensorParallelShardingStrategy):
+    def shard_model(
+        self,
+        model: nn.Module,
+        timeout_seconds: float,
+        on_timeout: TimeoutCallback | None,
+    ) -> nn.Module:
+        model = cast(LlamaModel, model)
+        for layer in model.layers:
+            # Force load weights before sharding to avoid FAST_SYNCH deadlock
+            eval_with_timeout(
+                layer.parameters(), timeout_seconds / len(model.layers), on_timeout
+            )
+            layer.self_attn.q_proj = self.all_to_sharded_linear(layer.self_attn.q_proj)
+            layer.self_attn.k_proj = self.all_to_sharded_linear(layer.self_attn.k_proj)
+            layer.self_attn.v_proj = self.all_to_sharded_linear(layer.self_attn.v_proj)
+            layer.self_attn.o_proj = self.sharded_to_all_linear(layer.self_attn.o_proj)
+            layer.self_attn.n_heads //= self.N
+            if layer.self_attn.n_kv_heads is not None:
+                layer.self_attn.n_kv_heads //= self.N
+
+            layer.mlp.gate_proj = self.all_to_sharded_linear(layer.mlp.gate_proj)
+            layer.mlp.down_proj = self.sharded_to_all_linear(layer.mlp.down_proj)
+            layer.mlp.up_proj = self.all_to_sharded_linear(layer.mlp.up_proj)
+
+        return model
+
+
+def _set_layers(model: nn.Module, layers: list[nn.Module]) -> None:
     inner_model_instance = _inner_model(model)
     if hasattr(inner_model_instance, "layers"):
         inner_model_instance.layers = layers
@@ -403,6 +453,105 @@ def _set_layers(model: nn.Module, layers: list[_LayerCallable]) -> None:
         raise ValueError("Model must have either a 'layers' or 'h' attribute")
 
 
+class DeepSeekShardingStrategy(TensorParallelShardingStrategy):
+    def shard_model(
+        self,
+        model: nn.Module,
+        timeout_seconds: float,
+        on_timeout: TimeoutCallback | None,
+    ) -> nn.Module:
+        model = cast(DeepseekV3Model, model)
+        for layer in model.layers:
+            eval_with_timeout(
+                layer.parameters(), timeout_seconds / len(model.layers), on_timeout
+            )
+            # Shard the self attention
+            if layer.self_attn.q_lora_rank is None:
+                layer.self_attn.q_proj = self.all_to_sharded_linear(
+                    layer.self_attn.q_proj
+                )
+            else:
+                layer.self_attn.q_b_proj = self.all_to_sharded_linear(
+                    layer.self_attn.q_b_proj
+                )
+            layer.self_attn.kv_b_proj = self.all_to_sharded_linear(
+                layer.self_attn.kv_b_proj
+            )
+            layer.self_attn.o_proj = self.sharded_to_all_linear(layer.self_attn.o_proj)
+            layer.self_attn.num_heads //= self.N
+
+            # Shard the MLP
+            if isinstance(layer.mlp, (DeepseekV3MLP, DeepseekV32MLP)):
+                layer.mlp.gate_proj = self.all_to_sharded_linear(layer.mlp.gate_proj)
+                layer.mlp.down_proj = self.sharded_to_all_linear(layer.mlp.down_proj)
+                layer.mlp.up_proj = self.all_to_sharded_linear(layer.mlp.up_proj)
+
+            # Shard the MoE. Shard in place since the MoE should be responsible
+            # for aggregating the results.
+            else:
+                self.all_to_sharded_linear_in_place(layer.mlp.shared_experts.gate_proj)
+                self.sharded_to_all_linear_in_place(layer.mlp.shared_experts.down_proj)
+                self.all_to_sharded_linear_in_place(layer.mlp.shared_experts.up_proj)
+                self.all_to_sharded_linear_in_place(layer.mlp.switch_mlp.gate_proj)
+                self.sharded_to_all_linear_in_place(layer.mlp.switch_mlp.down_proj)
+                self.all_to_sharded_linear_in_place(layer.mlp.switch_mlp.up_proj)
+                layer.mlp = ShardedDeepseekV3MoE(layer.mlp)  # type: ignore
+                layer.mlp.sharding_group = self.group
+
+        return model
+
+
+class ShardedDeepseekV3MoE(CustomMlxLayer):
+    def __init__(self, layer: nn.Module):
+        super().__init__(layer)
+        self.sharding_group: mx.distributed.Group | None = None
+
+    def __call__(self, x: mx.array) -> mx.array:
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+        y = self.original_layer.__call__(x) # type: ignore
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group)  # type: ignore
+        return y # type: ignore
+
+
+class MiniMaxShardingStrategy(TensorParallelShardingStrategy):
+    def shard_model(
+        self,
+        model: nn.Module,
+        timeout_seconds: float,
+        on_timeout: TimeoutCallback | None,
+    ) -> nn.Module:
+        model = cast(MiniMaxModel, model)
+        for layer in model.layers:
+            eval_with_timeout(
+                layer.parameters(), timeout_seconds / len(model.layers), on_timeout
+            )
+            # Shard the self attention
+            layer.self_attn.q_proj = self.all_to_sharded_linear(layer.self_attn.q_proj)
+            layer.self_attn.k_proj = self.all_to_sharded_linear(layer.self_attn.k_proj)
+            layer.self_attn.v_proj = self.all_to_sharded_linear(layer.self_attn.v_proj)
+            layer.self_attn.o_proj = self.sharded_to_all_linear(layer.self_attn.o_proj)
+            layer.self_attn.num_attention_heads //= self.N
+            layer.self_attn.num_key_value_heads //= self.N
+
+            # Shard the MoE. Shard in place since the MoE should be responsible
+            # for aggregating the results.
+            self.all_to_sharded_linear_in_place(
+                layer.block_sparse_moe.switch_mlp.gate_proj
+            )
+            self.sharded_to_all_linear_in_place(
+                layer.block_sparse_moe.switch_mlp.down_proj
+            )
+            self.all_to_sharded_linear_in_place(
+                layer.block_sparse_moe.switch_mlp.up_proj
+            )
+            layer.block_sparse_moe = ShardedQwenMoE(layer.block_sparse_moe)  # pyright: ignore[reportAttributeAccessIssue]
+            layer.block_sparse_moe.sharding_group = self.group  # pyright: ignore[reportAttributeAccessIssue]
+
+        return model
+
+
 class QwenShardingStrategy(TensorParallelShardingStrategy):
     def shard_model(
         self,
@@ -431,7 +580,7 @@ class QwenShardingStrategy(TensorParallelShardingStrategy):
                 self.all_to_sharded_linear_in_place(layer.mlp.switch_mlp.gate_proj)
                 self.sharded_to_all_linear_in_place(layer.mlp.switch_mlp.down_proj)
                 self.all_to_sharded_linear_in_place(layer.mlp.switch_mlp.up_proj)
-                layer.mlp = ShardedQwenMoE(layer.mlp)  # pyright: ignore[reportAttributeAccessIssue, reportArgumentType]
+                layer.mlp = ShardedQwenMoE(layer.mlp)  # pyright: ignore[reportAttributeAccessIssue]
                 layer.mlp.sharding_group = self.group
 
             # Shard the MLP
@@ -444,14 +593,69 @@ class QwenShardingStrategy(TensorParallelShardingStrategy):
 
 
 class ShardedQwenMoE(CustomMlxLayer):
-    def __init__(self, layer: _LayerCallable):
+    def __init__(self, layer: nn.Module):
         super().__init__(layer)
         self.sharding_group: mx.distributed.Group | None = None
 
     def __call__(self, x: mx.array) -> mx.array:
         if self.sharding_group is not None:
             x = sum_gradients(self.sharding_group)(x)
-        y = self.original_layer.__call__(x)
+        y = self.original_layer.__call__(x) # type: ignore
         if self.sharding_group is not None:
-            y = mx.distributed.all_sum(y, group=self.sharding_group)
-        return y
+            y = mx.distributed.all_sum(y, group=self.sharding_group) # type: ignore
+        return y # type: ignore
+
+
+class GptOssShardingStrategy(TensorParallelShardingStrategy):
+    def shard_model(
+        self,
+        model: nn.Module,
+        timeout_seconds: float,
+        on_timeout: TimeoutCallback | None,
+    ) -> nn.Module:
+        model = cast(GptOssMoeModel, model)
+
+        for layer in model.layers:
+            eval_with_timeout(
+                layer.parameters(), timeout_seconds / len(model.layers), on_timeout
+            )
+            layer.self_attn.q_proj = self.all_to_sharded_linear(layer.self_attn.q_proj)
+            layer.self_attn.k_proj = self.all_to_sharded_linear(layer.self_attn.k_proj)
+            layer.self_attn.v_proj = self.all_to_sharded_linear(layer.self_attn.v_proj)
+            layer.self_attn.o_proj = self.sharded_to_all_linear(layer.self_attn.o_proj)
+
+            layer.self_attn.num_attention_heads //= self.N
+            layer.self_attn.num_key_value_heads //= self.N
+            layer.self_attn.num_key_value_groups = (
+                layer.self_attn.num_attention_heads
+                // layer.self_attn.num_key_value_heads
+            )
+
+            layer.self_attn.sinks = layer.self_attn.sinks[
+                layer.self_attn.num_attention_heads
+                * self.group.rank() : layer.self_attn.num_attention_heads
+                * (self.group.rank() + 1)
+            ]
+
+            self.all_to_sharded_linear_in_place(layer.mlp.experts.gate_proj)
+            self.sharded_to_all_linear_in_place(layer.mlp.experts.down_proj)
+            self.all_to_sharded_linear_in_place(layer.mlp.experts.up_proj)
+
+            layer.mlp = ShardedGptOssMoE(layer.mlp)  # type: ignore
+            layer.mlp.sharding_group = self.group  # pyright: ignore[reportAttributeAccessIssue]
+
+        return model
+
+
+class ShardedGptOssMoE(CustomMlxLayer):
+    def __init__(self, layer: nn.Module):
+        super().__init__(layer)
+        self.sharding_group: mx.distributed.Group | None = None
+
+    def __call__(self, x: mx.array) -> mx.array:
+        if self.sharding_group is not None:
+            x = sum_gradients(self.sharding_group)(x)
+        y = self.original_layer(x) # type: ignore
+        if self.sharding_group is not None:
+            y = mx.distributed.all_sum(y, group=self.sharding_group) # type: ignore
+        return y # type: ignore