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

mlx_vlm/models/florence2/language.py

@@ -0,0 +1,452 @@
+import math
+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 SimpleKVCache
+from .config import TextConfig
+
+
+class Florence2Attention(nn.Module):
+    def __init__(
+        self, config: TextConfig, is_decoder: bool = False, is_causal: bool = False
+    ):
+        super().__init__()
+        self.embed_dim = config.d_model
+        self.num_heads = (
+            config.decoder_attention_heads
+            if is_decoder
+            else config.encoder_attention_heads
+        )
+        self.is_decoder = is_decoder
+        self.is_causal = is_causal
+        self.head_dim = self.embed_dim // self.num_heads
+        self.scaling = self.head_dim**-0.5
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def __call__(
+        self,
+        hidden_states,
+        key_value_states=None,
+        cache: Optional[SimpleKVCache] = None,
+        attention_mask=None,
+        layer_head_mask=None,
+    ):
+        batch_size, tgt_len, _ = hidden_states.shape
+
+        q = (
+            self.q_proj(hidden_states)
+            .reshape(batch_size, tgt_len, self.num_heads, self.head_dim)
+            .transpose(0, 2, 1, 3)
+        )
+
+        is_cross_attention = key_value_states is not None
+
+        batch_size, tgt_len, _ = hidden_states.shape
+        src_len = (
+            key_value_states.shape[1]
+            if key_value_states is not None
+            else hidden_states.shape[1]
+        )
+
+        if (
+            is_cross_attention
+            and cache is not None
+            and cache.cache_length > 0
+            and cache.keys.shape[2] == key_value_states.shape[1]
+        ):
+            # Cross-attention with cached keys/values - reuse them
+            k = cache.keys
+            v = cache.values
+
+        elif is_cross_attention:
+            # Cross attention - compute and cache keys/values from encoder
+            k = (
+                self.k_proj(key_value_states)
+                .reshape(batch_size, src_len, self.num_heads, self.head_dim)
+                .transpose(0, 2, 1, 3)
+            )
+            v = (
+                self.v_proj(key_value_states)
+                .reshape(batch_size, src_len, self.num_heads, self.head_dim)
+                .transpose(0, 2, 1, 3)
+            )
+            # Cache the cross-attention keys/values
+            if cache is not None:
+                cache.update(k, v)
+
+        elif cache is not None:
+            # Self-attention with cache - compute new k,v and concatenate with cache
+            k = (
+                self.k_proj(hidden_states)
+                .reshape(batch_size, src_len, self.num_heads, -1)
+                .transpose(0, 2, 1, 3)
+            )
+            v = (
+                self.v_proj(hidden_states)
+                .reshape(batch_size, src_len, self.num_heads, -1)
+                .transpose(0, 2, 1, 3)
+            )
+            # update_and_fetch handles cache concatenation
+            k, v = cache.update_and_fetch(k, v)
+
+        else:
+            # Self attention without cache (encoder)
+            k = (
+                self.k_proj(hidden_states)
+                .reshape(batch_size, src_len, self.num_heads, self.head_dim)
+                .transpose(0, 2, 1, 3)
+            )
+            v = (
+                self.v_proj(hidden_states)
+                .reshape(batch_size, src_len, self.num_heads, self.head_dim)
+                .transpose(0, 2, 1, 3)
+            )
+
+        if self.is_causal and self.is_decoder:
+            causal_mask = create_attention_mask(hidden_states)
+            attention_mask = causal_mask
+
+        attn_output = (
+            scaled_dot_product_attention(
+                q, k, v, cache=cache, scale=self.scaling, mask=attention_mask
+            )
+            .transpose(0, 2, 1, 3)
+            .reshape(batch_size, tgt_len, -1)
+        )
+
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output
+
+
+class Florence2EncoderLayer(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+        self.self_attn = Florence2Attention(config, is_decoder=False, is_causal=False)
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.activation_fn = nn.GELU()
+        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def __call__(self, hidden_states, attention_mask=None):
+        residual = hidden_states
+        hidden_states = self.self_attn(hidden_states, attention_mask=attention_mask)
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        return hidden_states
+
+
+class Florence2DecoderLayer(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.embed_dim = config.d_model
+        self.self_attn = Florence2Attention(config, is_decoder=True, is_causal=True)
+        self.dropout = config.dropout
+        self.activation_fn = nn.GELU()
+        self.activation_dropout = config.activation_dropout
+
+        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.encoder_attn = Florence2Attention(config, is_decoder=True)
+        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
+        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
+        self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+    def __call__(
+        self,
+        hidden_states,
+        encoder_hidden_states,
+        attention_mask=None,
+        encoder_attention_mask=None,
+        cache: Optional[Tuple[SimpleKVCache, SimpleKVCache]] = None,
+    ):
+        residual = hidden_states
+
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_cache = cache[0] if cache[0] is not None else None
+
+        hidden_states = self.self_attn(
+            hidden_states, attention_mask=attention_mask, cache=self_attn_cache
+        )
+
+        hidden_states = residual + hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+
+            # cross_attn cached key/values tuple is at positions 3,4 of cache tuple
+            cross_attn_cache = cache[-1] if cache[-1] is not None else None
+
+            hidden_states = self.encoder_attn(
+                hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                cache=cross_attn_cache,
+            )
+            hidden_states = residual + hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = residual + hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+
+        return hidden_states
+
+
+class Florence2Encoder(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.dropout = config.dropout
+        self.layerdrop = config.encoder_layerdrop
+
+        embed_dim = config.d_model
+        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
+        self.offset = 2
+        self.embed_positions = nn.Embedding(
+            config.max_position_embeddings + self.offset, embed_dim
+        )
+        self.layers = [
+            Florence2EncoderLayer(config) for _ in range(config.encoder_layers)
+        ]
+        self.layernorm_embedding = nn.LayerNorm(embed_dim)
+
+    def __call__(self, input_ids=None, inputs_embeds=None, attention_mask=None):
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+            input_shape = inputs_embeds.shape
+        else:
+            input_shape = inputs_embeds.shape
+
+        positions = mx.arange(input_shape[1])
+
+        if positions.ndim == 1:
+            positions = mx.expand_dims(positions, axis=0)
+
+        embed_pos = self.embed_positions(positions + self.offset)
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = self.layernorm_embedding(hidden_states)
+
+        for encoder_layer in self.layers:
+            # Add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = mx.random.uniform()
+            if self.training and (dropout_probability < self.layerdrop):
+                continue
+            hidden_states = encoder_layer(hidden_states, attention_mask)
+
+        return hidden_states
+
+
+class Florence2Decoder(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.dropout = config.dropout
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_position_embeddings
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+        self.offset = 2
+        self.embed_positions = nn.Embedding(
+            config.max_position_embeddings + self.offset, config.d_model
+        )
+        self.layers = [
+            Florence2DecoderLayer(config) for _ in range(config.decoder_layers)
+        ]
+        self.layernorm_embedding = nn.LayerNorm(config.d_model)
+
+    def __call__(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        head_mask=None,
+        cross_attn_head_mask=None,
+        inputs_embeds=None,
+        cache=None,
+    ):
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError(
+                "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            inputs_embeds = self.embed_tokens(input_ids)
+            input_shape = inputs_embeds.shape  # for 2d masks
+            positions = input_ids
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.shape[:-1]  # for 4d masks
+            positions = inputs_embeds[:, :, -1]
+        else:
+            raise ValueError(
+                "You have to specify either decoder_input_ids or decoder_inputs_embeds"
+            )
+
+        if positions.ndim == 1:
+            positions = mx.expand_dims(positions, axis=0)
+
+        cache_length = cache[0][0].keys.shape[2] if cache[0][0].cache_length > 0 else 0
+
+        bsz, seq_len = inputs_embeds.shape[:2]
+        positions = mx.arange(
+            cache_length,
+            cache_length + seq_len,
+            dtype=mx.int64,
+        )
+        positions = mx.expand_dims(positions, axis=0)
+
+        embed_pos = self.embed_positions(positions + self.offset)
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = self.layernorm_embedding(hidden_states)
+
+        for decoder_layer, c in zip(self.layers, cache):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = mx.random.uniform()
+            if self.training and (dropout_probability < self.layerdrop):
+                continue
+            hidden_states = decoder_layer(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=attention_mask,
+                encoder_attention_mask=encoder_attention_mask,
+                cache=c,
+            )
+
+        return hidden_states
+
+
+class Florence2LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.shared = nn.Embedding(config.vocab_size, config.d_model)
+        self.encoder = Florence2Encoder(config)
+        self.decoder = Florence2Decoder(config)
+        if config.scale_embedding:
+            self.embed_scale = math.sqrt(config.d_model)
+        else:
+            self.embed_scale = 1.0
+
+    def __call__(
+        self,
+        input_ids=None,
+        inputs_embeds=None,
+        decoder_input_ids=None,
+        decoder_inputs_embeds=None,
+        attention_mask=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        cache=None,
+    ):
+        self.encoder.embed_tokens = self.shared
+        self.decoder.embed_tokens = self.shared
+
+        if decoder_input_ids is None and decoder_inputs_embeds is None:
+            if input_ids is None:
+                raise ValueError(
+                    "If no `decoder_input_ids` or `decoder_inputs_embeds` are "
+                    "passed, `input_ids` cannot be `None`. Please pass either "
+                    "`input_ids` or `decoder_input_ids` or `decoder_inputs_embeds`."
+                )
+
+            decoder_input_ids = mx.zeros_like(input_ids)
+            decoder_input_ids[:, 1:] = input_ids[:, :-1]
+            decoder_input_ids[:, 0] = self.config.bos_token_id
+
+        if inputs_embeds is not None:
+            inputs_embeds = inputs_embeds * self.embed_scale
+
+        if cache is None:
+            cache = [(SimpleKVCache(), SimpleKVCache())] * len(self.decoder.layers)
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                inputs_embeds=inputs_embeds,
+                attention_mask=attention_mask,
+            )
+
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_outputs,
+            encoder_attention_mask=attention_mask,
+            inputs_embeds=decoder_inputs_embeds,
+            cache=cache,
+        )
+        return decoder_outputs, encoder_outputs
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.model = Florence2LanguageModel(config)
+        self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs=None,
+        inputs_embeds=None,
+        decoder_input_ids=None,
+        decoder_inputs_embeds=None,
+        attention_mask=None,
+        decoder_attention_mask=None,
+        encoder_outputs=None,
+        cache=None,
+        **kwargs,
+    ):
+        decoder_outputs, encoder_outputs = self.model(
+            inputs,
+            inputs_embeds,
+            decoder_input_ids,
+            decoder_inputs_embeds,
+            attention_mask,
+            decoder_attention_mask,
+            encoder_outputs,
+            cache,
+        )
+        out = self.lm_head(decoder_outputs)
+        return LanguageModelOutput(logits=out, encoder_outputs=encoder_outputs)
+
+    @property
+    def layers(self):
+        return range(self.model.config.decoder_layers)
+
+    @property
+    def head_dim(self):
+        return self.config.d_model // self.config.decoder_attention_heads
+
+    @property
+    def n_kv_heads(self):
+        return self.config.decoder_attention_heads
+
+    def make_cache(self):
+        return [(SimpleKVCache(), SimpleKVCache()) for n in self.layers]

mlx_vlm/models/florence2/processing_florence2.py

@@ -0,0 +1,30 @@
+from transformers.models.florence2.processing_florence2 import Florence2Processor
+
+# Store the original __init__
+_original_init = Florence2Processor.__init__
+
+
+def _patched_init(self, image_processor=None, tokenizer=None, **kwargs):
+    """Patched __init__ that adds image_token attributes to tokenizer if missing."""
+    if tokenizer is not None:
+        # Ensure tokenizer has image_token attribute
+        if not hasattr(tokenizer, "image_token"):
+            tokenizer.image_token = "<image>"
+
+        # Ensure tokenizer has image_token_id attribute
+        if not hasattr(tokenizer, "image_token_id"):
+            vocab = tokenizer.get_vocab()
+            if tokenizer.image_token in vocab:
+                tokenizer.image_token_id = vocab[tokenizer.image_token]
+            else:
+                tokenizer.add_tokens([tokenizer.image_token], special_tokens=True)
+                tokenizer.image_token_id = tokenizer.convert_tokens_to_ids(
+                    tokenizer.image_token
+                )
+
+    # Call original __init__
+    _original_init(self, image_processor=image_processor, tokenizer=tokenizer, **kwargs)
+
+
+# Apply the patch
+Florence2Processor.__init__ = _patched_init