optimize llama3.2 vision again (#12211)

python/llm/src/ipex_llm/transformers/convert.py

@@ -1023,6 +1023,9 @@ def _optimize_pre(model, qtype=None):
     if model.config.model_type == "llama":
         from ipex_llm.transformers.models.llama import merge_qkv
         model.apply(merge_qkv)
+    if model.config.model_type == "mllama":
+        from ipex_llm.transformers.models.mllama import merge_qkv
+        model.apply(merge_qkv)
     if model.config.model_type == "minicpm":
         from ipex_llm.transformers.models.minicpm import merge_qkv
         model.apply(merge_qkv)
@@ -1284,12 +1287,19 @@ def _optimize_post(model, lightweight_bmm=False):
         # llama 3.2 vision
         modeling_module_name = model.__class__.__module__
         module = importlib.import_module(modeling_module_name)
-        from ipex_llm.transformers.models.common import rms_norm_forward
-        from ipex_llm.transformers.models.common import mlp_silu_forward
         from ipex_llm.transformers.models.mllama import mllama_vision_attention_forward
         convert_forward(model, module.MllamaVisionAttention, mllama_vision_attention_forward)
+
+        from ipex_llm.transformers.models.common import rms_norm_forward
+        from ipex_llm.transformers.models.common import mlp_silu_forward
+        from ipex_llm.transformers.models.llama32 import llama_attention_forward
+        from ipex_llm.transformers.models.mllama import mllama_text_model_forward
+        from ipex_llm.transformers.models.mllama import mllama_cross_attention_forward
         convert_forward(model, module.MllamaTextRMSNorm, rms_norm_forward)
         convert_forward(model, module.MllamaTextMLP, mlp_silu_forward)
+        convert_forward(model, module.MllamaTextModel, mllama_text_model_forward)
+        convert_forward(model, module.MllamaTextSelfAttention, llama_attention_forward)
+        convert_forward(model, module.MllamaTextCrossAttention, mllama_cross_attention_forward)
     elif model.config.model_type == "llama":
         from transformers.models.llama.modeling_llama import LlamaRMSNorm
         from transformers.models.llama.modeling_llama import LlamaMLP

python/llm/src/ipex_llm/transformers/models/mllama.py

@@ -35,8 +35,23 @@
 import math
 import torch
 
-from typing import Optional
+from typing import Optional, Tuple, Union
-from ipex_llm.transformers.models.utils import use_sdp_non_causal
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.models.mllama.modeling_mllama import MllamaVisionAttention
+from transformers.models.mllama.modeling_mllama import MllamaTextSelfAttention
+from transformers.models.mllama.modeling_mllama import repeat_kv
+from ipex_llm.transformers.models.utils import use_sdp, use_sdp_causal, use_sdp_non_causal
+from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp8_kv_cache
+from ipex_llm.transformers.models.utils import should_use_fuse_rope
+from ipex_llm.transformers.models.common import merge_qkv_base, attention_softmax
+from ipex_llm.transformers.kv import DynamicNormalCache, DynamicFp8Cache
+from ipex_llm.transformers.utils import invalidInputError
+
+
+def merge_qkv(module: torch.nn.Module):
+    merge_qkv_base(module, MllamaVisionAttention)
+    merge_qkv_base(module, MllamaTextSelfAttention)
 
 
 def mllama_vision_attention_forward(
@@ -45,16 +60,12 @@ def mllama_vision_attention_forward(
     attention_mask: Optional[torch.Tensor] = None,
     output_attentions: bool = None,
 ):
-    query = self.q_proj(hidden_state)
+    bsz, q_len, _ = hidden_state.size()
-    key = self.k_proj(hidden_state)
-    value = self.v_proj(hidden_state)
 
-    batch_size, q_seq_len, _ = query.shape
+    qkv = self.qkv_proj(hidden_state)
-    _, kv_seq_len, _ = key.shape
+    qkv = qkv.view(bsz, q_len, 3 * self.num_heads, self.head_dim)
-
+    qkv = qkv.transpose(1, 2)
-    query = query.view(batch_size, q_seq_len, self.num_heads, self.head_dim).transpose(1, 2)
+    query, key, value = qkv.chunk(3, dim=1)
-    key = key.view(batch_size, kv_seq_len, self.num_heads, self.head_dim).transpose(1, 2)
-    value = value.view(batch_size, kv_seq_len, self.num_heads, self.head_dim).transpose(1, 2)
 
     if attention_mask is not None:  # no matter the length, we just slice it
         causal_mask = attention_mask[:, :, :, : key.shape[-2]]
@@ -79,7 +90,7 @@ def mllama_vision_attention_forward(
         attn_output = torch.matmul(attn_weights, value)
 
     attn_output = attn_output.transpose(1, 2).contiguous()
-    attn_output = attn_output.reshape(batch_size, q_seq_len, -1)
+    attn_output = attn_output.reshape(bsz, q_len, -1)
 
     output = self.o_proj(attn_output)
 
@@ -87,3 +98,227 @@ def mllama_vision_attention_forward(
         attn_weights = None
 
     return output, attn_weights
+
+
+def mllama_text_model_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    cross_attention_states: Optional[torch.FloatTensor] = None,
+    cross_attention_mask: Optional[torch.Tensor] = None,
+    full_text_row_masked_out_mask: Optional[torch.Tensor] = None,
+    past_key_values: Optional[Cache] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+) -> Union[Tuple, BaseModelOutputWithPast]:
+    output_attentions = (
+        output_attentions if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+    # IPEX-LLM OPT start: kv cache and quantize kv cache
+    inputs = input_ids if input_ids is not None else inputs_embeds
+    use_cache = True if inputs.device.type == "xpu" else use_cache
+    use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs)
+    if use_cache:
+        if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache):
+            past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)
+        elif not use_quantize_kv and not isinstance(past_key_values, DynamicNormalCache):
+            past_key_values = DynamicNormalCache.from_legacy_cache(past_key_values)
+    # IPEX-LLM OPT end
+
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    invalidInputError((input_ids is None) ^ (inputs_embeds is None),
+                      "You cannot specify both input_ids and inputs_embeds at the same time, "
+                      "and must specify either one")
+
+    if inputs_embeds is None:
+        inputs_embeds = self.embed_tokens(input_ids)
+
+    hidden_states = inputs_embeds
+
+    if cache_position is None:
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        cache_position = torch.arange(
+            past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1],
+            device=inputs_embeds.device
+        )
+    if position_ids is None:
+        position_ids = cache_position.unsqueeze(0)
+
+    causal_mask = self._update_causal_mask(
+        attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+    )
+
+    # create position embeddings to be shared across the decoder layers
+    position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+    # IPEX-LLM OPT start: use fused rope
+    if (should_use_fuse_rope(hidden_states, position_ids, False)
+            and self.rotary_emb.rope_type == "llama3"):
+        position_embeddings = self.rotary_emb.inv_freq
+    # IEPX_LLM OPT end
+
+    # decoder layers
+    all_hidden_states = () if output_hidden_states else None
+    all_self_attns = () if output_attentions else None
+    next_decoder_cache = None
+
+    for idx, decoder_layer in enumerate(self.layers):
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        # For text-only path we should skip cross attention layers.
+        # Let's check if the layer is cross attention layer and if we have cross attention states
+        # or cached cross attention states.
+        is_cross_attention_layer = idx in self.cross_attention_layers
+
+        # IPEX-LLM change start
+        if is_cross_attention_layer and cross_attention_states is None:
+            if past_key_values is None:
+                # use_cache=False
+                continue
+            elif len(past_key_values.key_cache) <= idx:
+                # first token but no cross_attention_states, means no image inputs
+                past_key_values.key_cache.append([])
+                past_key_values.value_cache.append([])
+                continue
+            elif past_key_values.key_cache[idx] == []:
+                # next token but no cross kv cache, means no image inputs
+                continue
+        # IPEX-LLM change end
+
+        layer_outputs = decoder_layer(
+            hidden_states,
+            cross_attention_states=cross_attention_states,
+            cross_attention_mask=cross_attention_mask,
+            attention_mask=causal_mask,
+            full_text_row_masked_out_mask=full_text_row_masked_out_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_values,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+        )
+
+        hidden_states = layer_outputs[0]
+
+        if use_cache:
+            next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+        if output_attentions:
+            all_self_attns += (layer_outputs[1],)
+
+    hidden_states = self.norm(hidden_states)
+
+    # add hidden states from the last decoder layer
+    if output_hidden_states:
+        all_hidden_states += (hidden_states,)
+
+    next_cache = next_decoder_cache if use_cache else None
+
+    if not return_dict:
+        return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
+                     if v is not None)
+    return BaseModelOutputWithPast(
+        last_hidden_state=hidden_states,
+        past_key_values=next_cache,
+        hidden_states=all_hidden_states,
+        attentions=all_self_attns,
+    )
+
+
+def mllama_cross_attention_forward(
+    self,
+    hidden_states: torch.Tensor,
+    cross_attention_states: Optional[torch.Tensor] = None,
+    past_key_value: Optional[Cache] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    output_attentions: bool = False,
+    use_cache: bool = None,
+    cache_position: Optional[torch.LongTensor] = None,
+):
+    bsz, q_len, _ = hidden_states.size()
+    query_states = self.q_proj(hidden_states)
+    query_states = self.q_norm(query_states.view(-1, self.head_dim))
+    query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+    if cross_attention_states is not None:
+        key_states = self.k_proj(cross_attention_states)
+        value_states = self.v_proj(cross_attention_states)
+        key_states = self.k_norm(key_states.view(-1, self.head_dim))
+        key_states = key_states.view(bsz, -1, self.num_key_value_heads,
+                                     self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, -1, self.num_key_value_heads,
+                                         self.head_dim).transpose(1, 2)
+
+        # if we have a new image + new tokens, we only computed key_states on that new image
+        # we still update the cross key states, past_image, new_image. And use it!
+        key_states, value_states = past_key_value.update(
+            key_states, value_states, self.layer_idx, None
+        )
+    else:
+        key_states, value_states = (
+            past_key_value.key_cache[self.layer_idx],
+            past_key_value.value_cache[self.layer_idx],
+        )
+
+    kv_seq_len = key_states.size(2)
+    if attention_mask is not None:  # no matter the length, we just slice it
+        causal_mask = attention_mask[:, :, :, :kv_seq_len]
+    else:
+        causal_mask = None
+
+    attn_weights = None
+    if use_sdp(q_len, kv_seq_len, self.head_dim, query_states):
+        import xe_addons
+        if isinstance(past_key_value, DynamicFp8Cache):
+            attn_output = xe_addons.sdp_fp8(query_states, key_states, value_states, causal_mask)
+        else:
+            attn_output = xe_addons.sdp(query_states, key_states, value_states, causal_mask)
+    elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training):
+        import xe_addons
+        if isinstance(past_key_value, DynamicFp8Cache):
+            attn_output = xe_addons.sdp_fp8_causal(query_states, key_states,
+                                                   value_states, causal_mask)
+        else:
+            attn_output = xe_addons.sdp_causal(query_states, key_states,
+                                               value_states, causal_mask)
+    else:
+        if isinstance(past_key_value, DynamicFp8Cache):
+            key_states, value_states = restore_fp8_kv_cache(key_states, value_states,
+                                                            query_states.dtype)
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states,
+                                    key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        if causal_mask is not None:
+            attn_weights = attn_weights + causal_mask
+
+        # upcast attention to fp32
+        attn_weights = attention_softmax(attn_weights, self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+    attn_output = attn_output.reshape(bsz, q_len, -1)
+    attn_output = self.o_proj(attn_output)
+
+    if not output_attentions:
+        attn_weights = None
+
+    return attn_output, attn_weights, past_key_value