PC: Add support for Starcoder2

llms/README.md

@@ -127,17 +127,18 @@ Most
 [Phi-2](https://huggingface.co/models?library=transformers,safetensors&other=phi&sort=trending),
 and
 [Mixtral](https://huggingface.co/models?library=transformers,safetensors&other=mixtral&sort=trending)
+[Starcoder2](https://huggingface.co/models?library=transformers,safetensors&other=starcoder2&sort=trending)
 style models should work out of the box.
 
 For some models (such as `Qwen` and `plamo`) the tokenizer requires you to
 enable the `trust_remote_code` option. You can do this by passing
 `--trust-remote-code` in the command line. If you don't specify the flag
 explicitly, you will be prompted to trust remote code in the terminal when
-running the model. 
+running the model.
 
 For `Qwen` models you must also specify the `eos_token`. You can do this by
 passing `--eos-token "<|endoftext|>"` in the command
-line. 
+line.
 
 These options can also be set in the Python API. For example:
 

llms/mlx_lm/models/starcoder2.py

@@ -0,0 +1,195 @@
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .base import BaseModelArgs
+from .layers import LayerNorm
+
+
+@dataclass
+class ModelArgs(BaseModelArgs):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    sliding_window: int
+    num_attention_heads: int
+    num_key_value_heads: int = None
+    vocab_size: int = 49152
+    intermediate_size: int = 12288
+    max_position_embeddings: int = 16384
+    norm_epsilon: float = 1e-05
+    rope_theta: float = 10000
+    rope_traditional: bool = False
+    attention_dropout: int = 0.0
+    residual_dropout: int = 0.0
+    embedding_dropout: int = 0.0
+    use_bias: bool = True
+    tie_word_embeddings: bool = True
+
+
+class Starcoder2Attention(nn.Module):
+    """
+    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
+    and "Generating Long Sequences with Sparse Transformers".
+    """
+
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.config = args
+        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
+        self.head_dim = head_dim = args.hidden_size // n_heads
+
+        self.repeats = n_heads // n_kv_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=True)
+
+        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 self.repeats > 1:
+            keys = mx.repeat(keys, self.repeats, axis=1)
+            values = mx.repeat(values, self.repeats, axis=1)
+
+        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)
+
+        scores = (queries * self.scale) @ keys.transpose(0, 1, 3, 2)
+        if mask is not None:
+            scores += mask
+        scores = mx.softmax(scores, axis=-1).astype(values.dtype)
+        output = (scores @ values).transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output), (keys, values)
+
+
+class Starcoder2MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.c_fc = nn.Linear(dim, hidden_dim, bias=True)
+        self.c_proj = nn.Linear(hidden_dim, dim, bias=True)
+
+    def __call__(self, x) -> mx.array:
+        return self.c_proj(nn.gelu(self.c_fc(x)))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: ModelArgs):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.hidden_size = args.hidden_size
+        self.self_attn = Starcoder2Attention(args)
+        self.mlp = Starcoder2MLP(args.hidden_size, args.intermediate_size)
+        self.input_layernorm = LayerNorm(args.hidden_size, eps=args.norm_epsilon)
+        self.post_attention_layernorm = LayerNorm(
+            args.hidden_size, eps=args.norm_epsilon
+        )
+        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 Starcoder2Model(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 = [
+            TransformerBlock(args=args) for _ in range(args.num_hidden_layers)
+        ]
+        self.norm = LayerNorm(args.hidden_size, eps=args.norm_epsilon)
+
+    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.model_type = args.model_type
+        self.model = Starcoder2Model(args)
+        self.tie_word_embeddings = args.tie_word_embeddings
+
+        # If tie_word_embeddings is False, tie (share) the embedding weights with lm_head
+        # Source: https://github.com/huggingface/transformers/blob/main/src/transformers/models/starcoder2/modeling_starcoder2.py#L1071
+        if not args.tie_word_embeddings:
+            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)
+
+        if not self.tie_word_embeddings:
+            return self.lm_head(out), cache
+        else:
+            out = out @ self.model.embed_tokens.weight.T
+            return out, cache
+
+    @property
+    def layers(self):
+        return self.model.layers

llms/mlx_lm/tuner/utils.py

@@ -32,6 +32,7 @@ def check_lora_layers(num_model):
         "stablelm",
         "qwen2",
         "gemma",
+        "starcoder2",
     ]:
         check_lora_layers(len(model.model.layers))