Add support for phi-3-vision-128k-instruct (#36)

* add phi3_v 

* Update test_models.py

* rebase branch

* remove debug print

* add prompt format

* add condition to fix quantisation

* bump version

---------

Co-authored-by: Prince Canuma <prince.gdt@gmail.com>

mlx_vlm/models/phi3_v/__init__.py

@@ -0,0 +1,8 @@
+from .phi3_v import (
+    LanguageModel,
+    Model,
+    ModelConfig,
+    TextConfig,
+    VisionConfig,
+    VisionModel,
+)

mlx_vlm/models/phi3_v/language.py

@@ -0,0 +1,19 @@
+import inspect
+from dataclasses import dataclass
+
+
+@dataclass
+class TextConfig:
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+
+class LanguageModel:
+    pass

mlx_vlm/models/phi3_v/phi3_v.py

@@ -0,0 +1,234 @@
+import inspect
+import math
+from dataclasses import dataclass
+from types import SimpleNamespace
+from typing import Dict, Optional, Tuple, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+from .language import LanguageModel, TextConfig
+from .su_rope import Phi3SuScaledRotaryEmbedding
+from .vision import VisionConfig, VisionModel
+
+
+@dataclass
+class ModelConfig:
+    text_config: TextConfig
+    vision_config: VisionConfig
+    model_type: str
+    vocab_size: int
+
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+
+    ignore_index: int = -100
+    image_token_index: int = 257152
+    hidden_size: int = 2048
+    pad_token_id: int = 0
+
+    num_key_value_heads: int = None
+    rope_theta: float = 10000
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    max_position_embeddings: int = 131072
+    original_max_position_embeddings: int = 4096
+
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+
+class Attention(nn.Module):
+    def __init__(self, args: TextConfig):
+        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
+        self.num_hidden_layers = args.num_hidden_layers
+
+        self.head_dim = head_dim = args.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        op_size = n_heads * head_dim + 2 * (n_kv_heads * head_dim)
+        self.qkv_proj = nn.Linear(dim, op_size, bias=False)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        rope_scale = 1.0
+        if args.rope_scaling and args.rope_scaling["type"] == "su":
+            self.rope = Phi3SuScaledRotaryEmbedding(
+                head_dim,
+                traditional=False,
+                base=args.rope_theta,
+                scale=rope_scale,
+                max_position_embeddings=args.max_position_embeddings,
+                original_max_position_embeddings=args.original_max_position_embeddings,
+                short_factor=args.rope_scaling["short_factor"],
+                long_factor=args.rope_scaling["long_factor"],
+            )
+        else:
+            if args.rope_scaling and args.rope_scaling["type"] == "linear":
+                rope_scale = 1 / args.rope_scaling["factor"]
+            self.rope = nn.RoPE(
+                head_dim,
+                traditional=args.rope_traditional,
+                base=args.rope_theta,
+                scale=rope_scale,
+            )
+
+    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
+
+        qkv = self.qkv_proj(x)
+        query_pos = self.n_heads * self.head_dim
+        queries, keys, values = mx.split(
+            qkv, [query_pos, query_pos + self.n_kv_heads * self.head_dim], axis=-1
+        )
+
+        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:
+            offset = cache[0].shape[2]
+            queries = self.rope(queries, offset=offset)
+            keys = self.rope(keys, offset=offset)
+            keys = mx.concatenate([cache[0], keys], axis=2)
+            values = mx.concatenate([cache[1], 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 MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_up_proj = nn.Linear(dim, 2 * hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        x = self.gate_up_proj(x)
+        gate, x = mx.split(x, 2, axis=-1)
+        return self.down_proj(nn.silu(gate) * x)
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, args: TextConfig):
+        super().__init__()
+        self.num_attention_heads = args.num_attention_heads
+        self.hidden_size = args.hidden_size
+        self.self_attn = Attention(args)
+        self.mlp = MLP(args.hidden_size, args.intermediate_size)
+        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 Phi3V(nn.Module):
+    def __init__(self, args: TextConfig):
+        super().__init__()
+        self.args = args
+        self.vocab_size = args.vocab_size
+        self.num_hidden_layers = args.num_hidden_layers
+        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
+        self.vision_embed_tokens = VisionModel(args)
+        self.layers = [
+            TransformerBlock(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,
+        pixel_values=None,
+        image_sizes=None,
+        cache=None,
+    ):
+        # print('inputs', inputs) # debug
+        h = self.embed_tokens(inputs)
+        p = np.argwhere(inputs < 0).tolist()
+        if pixel_values is not None:
+            h = self.vision_embed_tokens(pixel_values, h, image_sizes, p)
+        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 i, layer in enumerate(self.layers):
+            h, cache[i] = layer(h, mask, cache[i])
+        return self.norm(h), cache
+
+
+class Model(nn.Module):
+    def __init__(self, args: TextConfig):
+        super().__init__()
+        self.model_type = args.model_type
+        self.model = Phi3V(args)
+        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+        self.config = args
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        pixel_values=None,
+        mask=None,
+        cache=None,
+    ):
+        out, cache = self.model(inputs, pixel_values, mask, cache)
+        return self.lm_head(out).astype(self.lm_head.weight.dtype), cache
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def head_dim(self):
+        return self.args.hidden_size // self.args.num_attention_heads
+
+    @property
+    def n_kv_heads(self):
+        return self.args.num_key_value_heads
+
+    @property
+    def language_model(self):
+        return self
+
+    @property
+    def vision_model(self):
+        return self.model.vision_embed_tokens

mlx_vlm/models/phi3_v/su_rope.py

@@ -0,0 +1,71 @@
+import math
+
+import mlx.core as mx
+
+
+class Phi3SuScaledRotaryEmbedding:
+    def __init__(
+        self,
+        dims: int,
+        traditional: bool = False,
+        base: float = 10000.0,
+        scale: float = 1.0,
+        max_position_embeddings: int = 131072,
+        original_max_position_embeddings: int = 4096,
+        short_factor: list[float] | float = 1.0,
+        long_factor: list[float] | float = 1.0,
+    ):
+        """
+        Phi3Su Scaled Rotary Embedding layer for Phi-3 models.
+
+        Args:
+            dims (int): The feature dimensions to be rotated.
+            traditional (bool, optional): Unused. Default: ``False``.
+            base (int, optional): Base for the exponential scaling.
+            scale (float, optional): The scale used to scale the positions. Default: 1.0.
+            max_position_embeddings (int, optional): The maximum sequence length that this model was trained with. This is used to determine the size of the original RoPE embeddings when using long scaling. Default: 131072.
+            original_max_position_embeddings (int, optional): The maximum sequence length that this model was trained with. This is used to determine the size of the original RoPE embeddings when using long scaling. Default: 4096.
+            short_factor (float or list of floats, optional): List of scaling factors for sequences of length lesser than original_max_position_embeddings. Default: 1.0.
+            long_factor (float or list of floats, optional): List of scaling factors for sequences of length greater than original_max_position_embeddings.  Default: 1.0.
+        """
+        self.inv_freq_short = 1.0 / (
+            mx.array(short_factor, dtype=mx.float32)
+            * base ** (mx.arange(0, dims, 2, dtype=mx.float32) / dims)
+        )
+        self.inv_freq_long = 1.0 / (
+            scale
+            * mx.array(long_factor, dtype=mx.float32)
+            * base ** (mx.arange(0, dims, 2, dtype=mx.float32) / dims)
+        )
+        self.original_max_position_embeddings = original_max_position_embeddings
+        self.scaling_factor = math.sqrt(
+            1
+            + math.log(max_position_embeddings / original_max_position_embeddings)
+            / math.log(original_max_position_embeddings)
+        )
+
+    def _get_cos_sin(self, offset, L):
+        position_ids = mx.arange(offset, offset + L, dtype=mx.float32)[None]
+        inv_freq = (
+            self.inv_freq_long
+            if position_ids.max() + 1 > self.original_max_position_embeddings
+            else self.inv_freq_short
+        )
+        inv_freq_expanded = mx.repeat(
+            inv_freq[None, :, None], position_ids.shape[0], axis=0
+        )
+        position_ids_expanded = position_ids[:, None, :]
+        freqs = (inv_freq_expanded @ position_ids_expanded).transpose(0, 2, 1)
+        emb = mx.concatenate([freqs, freqs], axis=-1)
+        cos = mx.cos(emb) * self.scaling_factor
+        sin = mx.sin(emb) * self.scaling_factor
+        return mx.expand_dims(cos, axis=1), mx.expand_dims(sin, axis=1)
+
+    def __call__(self, x, offset: int = 0):
+        def _rotate_half(_x):
+            midpoint = _x.shape[-1] // 2
+            x1, x2 = _x[..., :midpoint], _x[..., midpoint:]
+            return mx.concatenate([-x2, x1], axis=-1)
+
+        cos, sin = self._get_cos_sin(offset, x.shape[2])
+        return (x * cos) + (_rotate_half(x) * sin)