Add support for qwen2moe (#640)

* add sparsemoe block and update decoder logic

* update file name to match HF

* update name

* Code formatting

* update gates calculation

* add support for Qwen2MoE.

* fix pytest

* code formatting and fix missing comma in utils

* Remove decoder sparse step.

Co-authored-by: bozheng-hit <dsoul0621@gmail.com>

* remove gate layer anti-quantisation

* remove unused argument

---------

Co-authored-by: bozheng-hit <dsoul0621@gmail.com>

llms/mlx_lm/models/qwen2_moe.py

@@ -0,0 +1,257 @@
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+from .base import BaseModelArgs
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    num_experts_per_tok: int
+    num_experts: int
+    moe_intermediate_size: int
+    shared_expert_intermediate_size: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int = None
+    rope_theta: float = 1000000
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = False
+
+    def __post_init__(self):
+        if self.num_key_value_heads is None:
+            self.num_key_value_heads = self.num_attention_heads
+
+        if self.rope_scaling:
+            required_keys = {"factor", "type"}
+            if not all(key in self.rope_scaling for key in required_keys):
+                raise ValueError(f"rope_scaling must contain keys {required_keys}")
+
+            if self.rope_scaling["type"] != "linear":
+                raise ValueError("rope_scaling 'type' currently only supports 'linear'")
+
+
+class Attention(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        head_dim = args.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=True)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=True)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=True)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        self.rope = nn.RoPE(
+            head_dim,
+            traditional=args.rope_traditional,
+            base=args.rope_theta,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if cache is not None:
+            key_cache, value_cache = cache
+            queries = self.rope(queries, offset=key_cache.shape[2])
+            keys = self.rope(keys, offset=key_cache.shape[2])
+            keys = mx.concatenate([key_cache, keys], axis=2)
+            values = mx.concatenate([value_cache, values], axis=2)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        output = mx.fast.scaled_dot_product_attention(
+            queries, keys, values, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output), (keys, values)
+
+
+class Qwen2MoeMLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class Qwen2MoeSparseMoeBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        dim = args.hidden_size
+        intermediate_size = args.moe_intermediate_size
+        shared_expert_intermediate_size = args.shared_expert_intermediate_size
+
+        self.num_experts = num_experts = args.num_experts
+        self.top_k = args.num_experts_per_tok
+
+        # gating
+        self.gate = nn.Linear(dim, num_experts, bias=False)
+        self.experts = [
+            Qwen2MoeMLP(dim, intermediate_size) for _ in range(self.num_experts)
+        ]
+        self.shared_expert = Qwen2MoeMLP(dim, shared_expert_intermediate_size)
+        self.shared_expert_gate = nn.Linear(dim, 1, bias=False)
+
+    def __call__(
+        self,
+        x: mx.array,
+    ):
+        ne = self.top_k
+        B, L, D = x.shape
+        x = x.reshape(-1, D)
+
+        # router_logits: (batch * sequence_length, n_experts)
+        gates = self.gate(x)
+        gates = mx.softmax(gates.astype(mx.float32), axis=-1)
+
+        inds = mx.stop_gradient(mx.argpartition(-gates, kth=ne, axis=-1)[:, :ne])
+
+        scores = mx.take_along_axis(gates, inds, axis=-1).astype(x.dtype)
+
+        if self.training:
+            inds = np.array(inds)
+            y = mx.zeros((B, ne, D), x.dtype)
+            for e, expert in enumerate(self.experts):
+                idx1, idx2 = map(mx.array, np.where(inds == e))
+                if idx1.size == 0:
+                    continue
+                y[idx1, idx2] = expert(x[idx1])
+
+            y = (y * scores[:, :, None]).sum(axis=1)
+        else:
+            y = []
+            for xt, st, it in zip(x, scores, inds.tolist()):
+                yt = mx.stack([self.experts[e](xt) for e in it], axis=-1)
+                yt = (yt * st).sum(axis=-1)
+                y.append(yt)
+
+            y = mx.stack(y, axis=0)
+
+        shared_expert_output = self.shared_expert(x)
+        shared_expert_output = (
+            mx.sigmoid(self.shared_expert_gate(x)) * shared_expert_output
+        )
+
+        y += shared_expert_output
+
+        return y.reshape(B, L, -1)
+
+
+class Qwen2MoeDecoderLayer(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.hidden_size = args.hidden_size
+        self.self_attn = Attention(args)
+        self.mlp = Qwen2MoeSparseMoeBlock(args)
+
+        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            args.hidden_size, eps=args.rms_norm_eps
+        )
+        self.args = args
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        r, cache = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out, cache
+
+
+class Qwen2MoeModel(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.layers = [
+            Qwen2MoeDecoderLayer(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        h = self.embed_tokens(inputs)
+
+        mask = None
+        if h.shape[1] > 1:
+            mask = nn.MultiHeadAttention.create_additive_causal_mask(h.shape[1])
+            mask = mask.astype(h.dtype)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for e, layer in enumerate(self.layers):
+            h, cache[e] = layer(h, mask, cache[e])
+
+        return self.norm(h), cache
+
+
+class Model(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.args = args
+        self.model_type = args.model_type
+        self.model = Qwen2MoeModel(args)
+        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+    ):
+        out, cache = self.model(inputs, cache)
+        return self.lm_head(out), cache
+
+    def sanitize(self, weights):
+        if self.args.tie_word_embeddings and "lm_head.weight" not in weights:
+            weights["lm_head.weight"] = weights["model.embed_tokens.weight"]
+        # Remove unused precomputed rotary freqs
+        return {
+            k: v for k, v in weights.items() if "self_attn.rotary_emb.inv_freq" not in k
+        }
+
+    @property
+    def layers(self):
+        return self.model.layers

llms/mlx_lm/tuner/utils.py

@@ -70,6 +70,7 @@ def linear_to_lora_layers(
         "mixtral",
         "stablelm",
         "qwen2",
+        "qwen2_moe",
         "gemma",
         "starcoder2",
         "cohere",
@@ -77,6 +78,9 @@ def linear_to_lora_layers(
         keys = set(["self_attn.q_proj", "self_attn.v_proj"])
         if model.model_type == "mixtral":
             keys.add("block_sparse_moe.gate")
+        if model.model_type == "qwen2_moe":
+            keys.add("mlp.gate")
+            keys.add("mlp.shared_expert_gate")
     elif model.model_type == "olmo":
         keys = set(["att_proj"])
     elif model.model_type == "phi-msft":

llms/tests/test_models.py

@@ -114,6 +114,27 @@ class TestModels(unittest.TestCase):
             args.n_layers,
         )
 
+    def test_qwen2_moe(self):
+        from mlx_lm.models import qwen2_moe
+
+        args = qwen2_moe.ModelArgs(
+            model_type="qwen2_moe",
+            hidden_size=1024,
+            num_hidden_layers=4,
+            intermediate_size=2048,
+            num_attention_heads=4,
+            rms_norm_eps=1e-5,
+            vocab_size=10_000,
+            num_experts_per_tok=4,
+            num_experts=16,
+            moe_intermediate_size=1024,
+            shared_expert_intermediate_size=2048,
+        )
+        model = qwen2_moe.Model(args)
+        self.model_test_runner(
+            model, args.model_type, args.vocab_size, args.num_hidden_layers
+        )
+
     def test_qwen2(self):
         from mlx_lm.models import qwen2