fount-vlm-nell-02 0.3.11__py3-none-any.whl

mlx_vlm/models/mllama/language.py

@@ -0,0 +1,377 @@
+from typing import Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from ..base import (
+    LanguageModelOutput,
+    create_attention_mask,
+    scaled_dot_product_attention,
+)
+from ..cache import KVCache
+from .config import TextConfig
+
+
+class MllamaTextCrossAttention(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.layer_idx = layer_idx
+        self.scale = self.head_dim**-0.5
+        self.q_proj = nn.Linear(
+            self.hidden_size, self.num_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            self.num_heads * self.head_dim, self.hidden_size, bias=False
+        )
+
+        self.q_norm = nn.RMSNorm(self.head_dim, eps=config.rms_norm_eps)
+        self.k_norm = nn.RMSNorm(self.head_dim, eps=config.rms_norm_eps)
+
+    def __call__(
+        self,
+        hidden_states: mx.array,
+        cross_attention_states: Optional[mx.array] = None,
+        attention_mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+    ) -> mx.array:
+
+        bsz, q_len, _ = hidden_states.shape
+        query = (
+            self.q_proj(hidden_states)
+            .reshape(bsz, q_len, self.num_heads, self.head_dim)
+            .transpose(0, 2, 1, 3)
+        )
+        query_states = self.q_norm(query)
+
+        if cross_attention_states is not None:
+            key_states = (
+                self.k_proj(cross_attention_states)
+                .reshape(bsz, -1, self.num_key_value_heads, self.head_dim)
+                .transpose(0, 2, 1, 3)
+            )
+            value_states = (
+                self.v_proj(cross_attention_states)
+                .reshape(bsz, -1, self.num_key_value_heads, self.head_dim)
+                .transpose(0, 2, 1, 3)
+            )
+            key_states = self.k_norm(key_states)
+        elif cache is not None and cache.offset > 0:
+            key_states, value_states = cache.fetch()
+        else:
+            key_states, value_states = mx.split(query, 2, axis=1)
+            key_states = self.k_norm(key_states)
+
+        attn_output = scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            cache,
+            scale=self.scale,
+            mask=attention_mask,  # add a dim for batch processing
+        )
+        attn_output = attn_output.transpose(0, 2, 1, 3).reshape(
+            bsz, q_len, self.hidden_size
+        )
+        return self.o_proj(attn_output)
+
+
+class MllamaTextSelfAttention(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.scale = self.head_dim**-0.5
+        self.layer_idx = layer_idx
+
+        self.q_proj = nn.Linear(
+            self.hidden_size, self.num_heads * self.head_dim, bias=False
+        )
+        self.k_proj = nn.Linear(
+            self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.v_proj = nn.Linear(
+            self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False
+        )
+        self.o_proj = nn.Linear(
+            self.num_heads * self.head_dim, self.hidden_size, bias=False
+        )
+
+        self.rope = nn.RoPE(
+            self.head_dim,
+            traditional=config.rope_traditional,
+            base=config.rope_theta,
+            scale=1,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+    ) -> mx.array:
+        bsz, q_len, _ = x.shape
+        query_states = (
+            self.q_proj(x).reshape(bsz, q_len, self.num_heads, -1).transpose(0, 2, 1, 3)
+        )
+        key_states = (
+            self.k_proj(x)
+            .reshape(bsz, q_len, self.num_key_value_heads, -1)
+            .transpose(0, 2, 1, 3)
+        )
+        value_states = (
+            self.v_proj(x)
+            .reshape(bsz, q_len, self.num_key_value_heads, -1)
+            .transpose(0, 2, 1, 3)
+        )
+
+        if cache is not None:
+            query_states = self.rope(query_states, offset=cache.offset)
+            key_states = self.rope(key_states, offset=cache.offset)
+            key_states, value_states = cache.update_and_fetch(key_states, value_states)
+        else:
+            query_states = self.rope(query_states)
+            key_states = self.rope(key_states)
+
+        attn_output = scaled_dot_product_attention(
+            query_states, key_states, value_states, cache, scale=self.scale, mask=mask
+        )
+        attn_output = attn_output.transpose(0, 2, 1, 3).reshape(
+            bsz, q_len, self.hidden_size
+        )
+        return self.o_proj(attn_output)
+
+
+class MllamaTextMLP(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.gate_proj = nn.Linear(
+            config.hidden_size, config.intermediate_size, bias=False
+        )
+        self.up_proj = nn.Linear(
+            config.hidden_size, config.intermediate_size, bias=False
+        )
+        self.down_proj = nn.Linear(
+            config.intermediate_size, config.hidden_size, bias=False
+        )
+        self.act_fn = lambda x: x * mx.sigmoid(x)
+
+    def __call__(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class MllamaSelfAttentionDecoderLayer(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.self_attn = MllamaTextSelfAttention(config, layer_idx=layer_idx)
+        self.mlp = MllamaTextMLP(config)
+        self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+
+    def __call__(
+        self,
+        hidden_states: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+    ) -> mx.array:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states = self.self_attn(
+            x=hidden_states,
+            mask=mask,
+            cache=cache,
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class MllamaCrossAttentionDecoderLayer(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.cross_attn = MllamaTextCrossAttention(config, layer_idx=layer_idx)
+        self.mlp = MllamaTextMLP(config)
+        self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.cross_attn_attn_gate = mx.zeros(1)
+        self.cross_attn_mlp_gate = mx.zeros(1)
+
+    def __call__(
+        self,
+        hidden_states: mx.array,
+        cross_attention_states: mx.array,
+        attention_mask: Optional[mx.array] = None,
+        full_text_row_masked_out_mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+    ) -> mx.array:
+        residual = hidden_states
+        hidden_states = self.input_layernorm(hidden_states)
+        hidden_states = self.cross_attn(
+            hidden_states=hidden_states,
+            cross_attention_states=cross_attention_states,
+            attention_mask=attention_mask,
+            cache=cache,
+        )
+        hidden_states = residual + mx.tanh(self.cross_attn_attn_gate) * hidden_states
+
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        if full_text_row_masked_out_mask is not None:
+            hidden_states = full_text_row_masked_out_mask[:, 0] * hidden_states
+        hidden_states = residual + mx.tanh(self.cross_attn_mlp_gate) * hidden_states
+
+        return hidden_states
+
+
+class MllamaTextModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.vocab_size = config.vocab_size
+        self.hidden_size = config.hidden_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size + 8, config.hidden_size)
+        self.layers = [
+            (
+                MllamaCrossAttentionDecoderLayer(config, layer_idx)
+                if layer_idx in config.cross_attention_layers
+                else MllamaSelfAttentionDecoderLayer(config, layer_idx)
+            )
+            for layer_idx in range(config.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def __call__(
+        self,
+        input_ids: Optional[mx.array] = None,
+        mask: Optional[mx.array] = None,
+        position_ids: Optional[mx.array] = None,
+        cross_attention_states: Optional[mx.array] = None,
+        cross_attention_mask: Optional[mx.array] = None,
+        full_text_row_masked_out_mask: Optional[mx.array] = None,
+        inputs_embeds: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+    ) -> mx.array:
+        # Prioritize inputs_embeds if provided
+        if inputs_embeds is not None:
+            batch_size, seq_length, _ = inputs_embeds.shape
+        elif input_ids is not None:
+            batch_size, seq_length = input_ids.shape
+            inputs_embeds = self.embed_tokens(input_ids)
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        if position_ids is None:
+            position_ids = mx.expand_dims(mx.arange(seq_length), 0)
+            position_ids = mx.repeat(position_ids, batch_size, axis=0)
+
+        hidden_states = inputs_embeds
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        if mask is None:
+            mask = create_attention_mask(hidden_states, cache)
+
+        for idx, (decoder_layer, c) in enumerate(zip(self.layers, cache)):
+            if idx in self.config.cross_attention_layers:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    cross_attention_states=cross_attention_states,
+                    attention_mask=cross_attention_mask,
+                    full_text_row_masked_out_mask=full_text_row_masked_out_mask,
+                    cache=c,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    mask=mask,
+                    cache=c,
+                )
+            hidden_states = layer_outputs
+
+        hidden_states = self.norm(hidden_states)
+
+        return hidden_states
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.model = MllamaTextModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: Optional[mx.array] = None,
+        inputs_embeds: Optional[mx.array] = None,
+        mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+        cross_attention_states: Optional[mx.array] = None,
+        cross_attention_mask: Optional[mx.array] = None,
+        full_text_row_masked_out_mask: Optional[mx.array] = None,
+        **kwargs,
+    ) -> Tuple[mx.array, Optional[mx.array]]:
+
+        hidden_states = self.model(
+            input_ids=inputs,
+            mask=mask,
+            cross_attention_states=cross_attention_states,
+            cross_attention_mask=cross_attention_mask,
+            full_text_row_masked_out_mask=full_text_row_masked_out_mask,
+            inputs_embeds=inputs_embeds,
+            cache=cache,
+        )
+
+        logits = self.lm_head(hidden_states)
+
+        return LanguageModelOutput(
+            logits=logits, cross_attention_states=cross_attention_states
+        )
+
+    @staticmethod
+    def sanitize(weights):
+        # 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
+
+    @property
+    def head_dim(self):
+        return self.config.hidden_size // self.config.num_attention_heads
+
+    @property
+    def n_kv_heads(self):
+        return self.config.num_key_value_heads

mlx_vlm/models/mllama/mllama.py

@@ -0,0 +1,210 @@
+from typing import Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from ..base import InputEmbeddingsFeatures
+from ..cache import KVCache
+from .config import ModelConfig
+from .language import LanguageModel
+from .vision import VisionModel
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.config = config
+        self.vision_tower = VisionModel(config.vision_config)
+        self.language_model = LanguageModel(config.text_config)
+        self.multi_modal_projector = nn.Linear(
+            config.vision_config.vision_output_dim,
+            config.text_config.hidden_size,
+            bias=True,
+        )
+
+    @property
+    def layers(self):
+        return self.language_model.model.layers
+
+    def get_input_embeddings(
+        self,
+        input_ids: Optional[mx.array] = None,
+        pixel_values: Optional[mx.array] = None,
+        **kwargs,
+    ):
+        aspect_ratio_ids = kwargs.get("aspect_ratio_ids", None)
+        aspect_ratio_mask = kwargs.get("aspect_ratio_mask", None)
+        cross_attention_mask = kwargs.get("cross_attention_mask", None)
+
+        # Get text embeddings
+        inputs_embeds = self.language_model.model.embed_tokens(input_ids)
+
+        cross_attention_states = None
+        full_text_row_masked_out_mask = None
+
+        # Process vision input if provided
+        if pixel_values is not None:
+            if aspect_ratio_ids is None:
+                raise ValueError(
+                    "`aspect_ratio_ids` must be provided if `pixel_values` is provided"
+                )
+
+            vision_outputs = self.vision_tower(
+                pixel_values=pixel_values,
+                aspect_ratio_ids=aspect_ratio_ids,
+                aspect_ratio_mask=aspect_ratio_mask,
+            )
+            cross_attention_states = vision_outputs[0]
+
+            cross_attention_states = self.multi_modal_projector(
+                cross_attention_states
+            ).reshape(
+                -1,
+                cross_attention_states.shape[-2],
+                self.config.text_config.hidden_size,
+            )
+
+        # Prepare cross attention mask
+        if cross_attention_mask is not None:
+            cross_attention_mask, full_text_row_masked_out_mask = (
+                self._prepare_cross_attention_mask(
+                    cross_attention_mask,
+                    num_vision_tokens=(
+                        self.config.vision_config.image_size
+                        // self.config.vision_config.patch_size
+                    )
+                    ** 2
+                    + 1,
+                )
+            )
+
+            cache_position = mx.arange(input_ids.shape[1], dtype=mx.int32)
+            cross_attention_mask = cross_attention_mask[:, :, cache_position]
+            full_text_row_masked_out_mask = full_text_row_masked_out_mask[
+                :, :, cache_position
+            ]
+
+        return InputEmbeddingsFeatures(
+            inputs_embeds=inputs_embeds,
+            cross_attention_states=cross_attention_states,
+            cross_attention_mask=cross_attention_mask,
+            full_text_row_masked_out_mask=full_text_row_masked_out_mask,
+        )
+
+    def __call__(
+        self,
+        input_ids: mx.array,
+        pixel_values: Optional[mx.array] = None,
+        mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+        **kwargs,
+    ) -> Tuple[mx.array, Optional[mx.array]]:
+
+        aspect_ratio_ids = kwargs.pop("aspect_ratio_ids", None)
+        aspect_ratio_mask = kwargs.pop("aspect_ratio_mask", None)
+        cross_attention_mask = kwargs.pop("cross_attention_mask", None)
+
+        inputs_embeds = None
+
+        # Process vision input if provided
+        if pixel_values is not None:
+            if aspect_ratio_ids is None:
+                raise ValueError(
+                    "`aspect_ratio_ids` must be provided if `pixel_values` is provided"
+                )
+
+            vision_outputs = self.vision_tower(
+                pixel_values=pixel_values,
+                aspect_ratio_ids=aspect_ratio_ids,
+                aspect_ratio_mask=aspect_ratio_mask,
+            )
+            cross_attention_states = vision_outputs[0]
+
+            cross_attention_states = self.multi_modal_projector(
+                cross_attention_states
+            ).reshape(
+                -1,
+                cross_attention_states.shape[-2],
+                self.config.text_config.hidden_size,
+            )
+
+        else:
+            cross_attention_states = None
+
+        # Prepare cross attention mask
+        if cross_attention_mask is not None:
+            cross_attention_mask, full_text_row_masked_out_mask = (
+                self._prepare_cross_attention_mask(
+                    cross_attention_mask,
+                    num_vision_tokens=(
+                        self.config.vision_config.image_size
+                        // self.config.vision_config.patch_size
+                    )
+                    ** 2
+                    + 1,
+                )
+            )
+        else:
+            full_text_row_masked_out_mask = None
+
+        if cross_attention_mask is not None:
+            cache_position = mx.arange(input_ids.shape[1], dtype=mx.int32)
+            cross_attention_mask = cross_attention_mask[:, :, cache_position]
+            full_text_row_masked_out_mask = full_text_row_masked_out_mask[
+                :, :, cache_position
+            ]
+
+        # Process language input
+        outputs = self.language_model(
+            inputs=input_ids,
+            mask=mask,
+            cross_attention_states=cross_attention_states,
+            cross_attention_mask=cross_attention_mask,
+            full_text_row_masked_out_mask=full_text_row_masked_out_mask,
+            inputs_embeds=inputs_embeds,
+            cache=cache,
+        )
+
+        return outputs
+
+    def _prepare_cross_attention_mask(
+        self,
+        cross_attention_mask: mx.array,
+        num_vision_tokens: int,
+    ) -> Tuple[mx.array, mx.array]:
+        batch_size, text_total_length, *_ = cross_attention_mask.shape
+        cross_attention_mask = mx.repeat(
+            cross_attention_mask, num_vision_tokens, axis=3
+        )
+        cross_attention_mask = cross_attention_mask.reshape(
+            batch_size, text_total_length, -1
+        )
+        cross_attention_mask = mx.expand_dims(cross_attention_mask, 1)
+
+        # Invert the mask
+        inverted_cross_attn_mask = 1.0 - cross_attention_mask
+        fill_array = mx.array(-1e9)
+        fill_array = mx.broadcast_to(fill_array, inverted_cross_attn_mask.shape)
+        cross_attention_mask = mx.where(
+            inverted_cross_attn_mask,
+            fill_array,
+            cross_attention_mask,
+        )
+
+        # Apply full-row bias
+        full_text_row_masked_out_mask = mx.any(
+            cross_attention_mask != -1e9,
+            axis=-1,
+            keepdims=True,
+        )
+        cross_attention_mask *= full_text_row_masked_out_mask
+
+        return cross_attention_mask, full_text_row_masked_out_mask
+
+    def sanitize(self, weights):
+        def transform_key(key):
+            if "vision_tower" not in key:
+                key = key.replace("vision_model", "vision_tower")
+            return key
+
+        return {transform_key(k): v for k, v in weights.items()}