add basic support and optimization for qwen2-vl (#12104)

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

@@ -1000,6 +1000,9 @@ def _optimize_pre(model, qtype=None):
     if model.config.model_type == "qwen2_audio":
         from ipex_llm.transformers.models.qwen2 import merge_qkv
         model.language_model.apply(merge_qkv)
+    if model.config.model_type == "qwen2_vl":
+        from ipex_llm.transformers.models.qwen2_vl import merge_qkv
+        model.apply(merge_qkv)
     if model.config.model_type == "stablelm":
         # For stablelm-zephyr-3b and stablelm-2-zephyr-1_6b
         from ipex_llm.transformers.models.stablelm import merge_qkv
@@ -1651,6 +1654,17 @@ def _optimize_post(model, lightweight_bmm=False):
                         qwen2_attention_forward)
     elif model.config.model_type == "qwen2_audio":
         _optimize_post(model.language_model, lightweight_bmm=lightweight_bmm)
+    elif model.config.model_type == "qwen2_vl":
+        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.qwen2 import qwen2_mlp_forward
+        from ipex_llm.transformers.models.qwen2_vl import qwen2_vl_model_forward
+        from ipex_llm.transformers.models.qwen2_vl import qwen2_vl_attention_forward
+        convert_forward(model, module.Qwen2RMSNorm, rms_norm_forward)
+        convert_forward(model, module.Qwen2MLP, qwen2_mlp_forward)
+        convert_forward(model, module.Qwen2VLModel, qwen2_vl_model_forward)
+        convert_forward(model, module.Qwen2VLAttention, qwen2_vl_attention_forward)
     elif model.config.model_type == "cohere":
         # for CohereForAI/c4ai-command-r-v01
         invalidInputError(version.parse(trans_version) >= version.parse("4.40.0"),

python/llm/src/ipex_llm/transformers/models/qwen2_vl.py

@@ -0,0 +1,182 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# Some parts of this file is adapted from
+# https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py
+# which is licensed under Apache License 2.0:
+#
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+import math
+from typing import Optional, Tuple, Union, List
+
+import torch
+
+from ipex_llm.transformers.models.common import merge_qkv_base
+from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp8_kv_cache
+from ipex_llm.transformers.models.utils import use_sdp, use_sdp_causal
+from ipex_llm.transformers.kv import DynamicFp8Cache, DynamicNormalCache
+
+from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLAttention, Qwen2VLModel
+from transformers.models.qwen2_vl.modeling_qwen2_vl import apply_multimodal_rotary_pos_emb
+from transformers.models.qwen2_vl.modeling_qwen2_vl import repeat_kv
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.cache_utils import Cache
+
+
+def merge_qkv(module: torch.nn.Module):
+    merge_qkv_base(module, Qwen2VLAttention)
+
+
+def qwen2_vl_model_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = 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]:
+    # IPEX-LLM OPT: kv cache and quantize kv cache and sdp
+    inputs = input_ids if input_ids is not None else inputs_embeds
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+    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)
+
+    return Qwen2VLModel.forward(
+        self=self,
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+    )
+
+
+def qwen2_vl_attention_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Cache] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+    cache_position: Optional[torch.LongTensor] = None,
+    position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]]=None,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    bsz, q_len, _ = hidden_states.size()
+
+    qkv = self.qkv_proj(hidden_states)
+    qkv = qkv.view(bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim)
+    qkv = qkv.transpose(1, 2)
+    query_states, key_states, value_states = qkv.split([self.num_heads,
+                                                        self.num_key_value_heads,
+                                                        self.num_key_value_heads], dim=1)
+
+    if position_embeddings is None:
+        cos, sin = self.rotary_emb(value_states, position_ids)
+    else:
+        cos, sin = position_embeddings
+    query_states, key_states = apply_multimodal_rotary_pos_emb(
+        query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
+    )
+
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is not None:
+        cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+        key_states, value_states = past_key_value.update(key_states, value_states,
+                                                         self.layer_idx, cache_kwargs)
+        kv_seq_len = key_states.shape[-2]
+
+    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,
+                                            attention_mask)
+        else:
+            attn_output = xe_addons.sdp(query_states, key_states, value_states,
+                                        attention_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, attention_mask)
+        else:
+            attn_output = xe_addons.sdp_causal(query_states, key_states,
+                                               value_states, attention_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 attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+
+        # upcast attention to fp32
+        attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1,
+                                                   dtype=torch.float32).to(query_states.dtype)
+        attn_weights = torch.nn.functional.dropout(attn_weights, p=self.attention_dropout,
+                                                   training=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