LLM: Optimize qwen1.5 moe model (#10706)

* update moe block

* fix style

* enable optmize MLP

* enabel kv_cache

* enable fuse rope

* enable fused qkv

* enable flash_attention

* error sdp quantize

* use old api

* use fuse

* use xetla

* fix python style

* update moe_blocks num

* fix output error

* add cpu sdpa

* update

* update

* update

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

@@ -1154,6 +1154,28 @@ def _optimize_post(model, lightweight_bmm=False):
         convert_forward(model,
                         module.Qwen2Attention,
                         qwen2_attention_forward)
+    elif model.config.model_type == "qwen2_moe":
+        # for Qwen1.5-MOE-A2.7B
+        modeling_module_name = model.__class__.__module__
+        module = importlib.import_module(modeling_module_name)
+        from ipex_llm.transformers.models.qwen2_moe import qwen2moe_moeblock_forward
+        from ipex_llm.transformers.models.qwen2_moe import qwen2moe_attention_forward
+        from ipex_llm.transformers.models.qwen2_moe import qwen2moe_model_forward
+        convert_forward(model,
+                        module.Qwen2MoeModel,
+                        qwen2moe_model_forward)
+        convert_forward(model,
+                        module.Qwen2MoeRMSNorm,
+                        llama_rms_norm_forward)
+        convert_forward(model,
+                        module.Qwen2MoeSparseMoeBlock,
+                        qwen2moe_moeblock_forward)
+        convert_forward(model,
+                        module.Qwen2MoeMLP,
+                        llama_mlp_forward)
+        convert_forward(model,
+                        module.Qwen2MoeAttention,
+                        qwen2moe_attention_forward)
     elif model.config.model_type == "aquila":
         modeling_module_name = model.__class__.__module__
         module = importlib.import_module(modeling_module_name)

python/llm/src/ipex_llm/transformers/models/qwen2_moe.py

@@ -0,0 +1,622 @@
+#
+# 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_moe/modeling_qwen2_moe.py
+
+# coding=utf-8
+# 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.
+
+""" PyTorch Qwen2MoE model."""
+import math
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+import torch.utils.checkpoint
+import warnings
+from typing import TYPE_CHECKING, Optional, Tuple, Union, Callable, List
+from ipex_llm.transformers.models.llama import repeat_kv
+from ipex_llm.transformers.models.qwen2 import should_use_fuse_rope
+from ipex_llm.transformers.models.utils import extend_kv_cache, append_kv_cache
+from ipex_llm.transformers.models.utils import apply_rotary_pos_emb_cache_freq_xpu
+from ipex_llm.transformers.models.utils import is_enough_kv_cache_room_4_36
+from transformers.models.qwen2.modeling_qwen2 import apply_rotary_pos_emb
+from ipex_llm.utils.common import invalidInputError
+from ipex_llm.transformers.models.utils import decoding_fast_path_qtype_check
+from ipex_llm.transformers.models.utils import use_flash_attention, use_esimd_sdp
+from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeModel, apply_rotary_pos_emb
+from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp8_kv_cache
+from ipex_llm.transformers.kv import DynamicFp8Cache
+
+import os
+
+KV_CACHE_ALLOC_BLOCK_LENGTH = int(os.environ.get("KV_CACHE_ALLOC_BLOCK_LENGTH", 256))
+
+
+def qwen2moe_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,
+    output_router_logits: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+):
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+    if use_cache and use_quantize_kv_cache(self.layers[0].mlp.shared_expert.up_proj, input_ids):
+        if not isinstance(past_key_values, DynamicFp8Cache):
+            past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)
+    return Qwen2MoeModel.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,
+        output_router_logits=output_router_logits,
+        return_dict=return_dict,
+    )
+
+
+def qwen2moe_attention_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+    **kwargs,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    if use_quantize_kv_cache(self.q_proj, hidden_states):
+        forward_function = qwen2moe_attention_forward_quantized
+    elif hidden_states.device.type == "cpu":
+        forward_function = qwen2moe_attention_forward_sdpa
+    else:
+        forward_function = qwen2moe_attention_forward_origin
+    return forward_function(
+        self=self,
+        hidden_states=hidden_states,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_value=past_key_value,
+        output_attentions=output_attentions,
+        use_cache=use_cache,
+        **kwargs,
+    )
+
+
+def qwen2moe_attention_forward_quantized(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+    **kwargs,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    if "padding_mask" in kwargs:
+        warnings.warn(
+            "Passing `padding_mask` is deprecated and will be removed in v4.37."
+            "Please make sure use `attention_mask` instead.`"
+        )
+    use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids)
+    bsz, q_len, _ = hidden_states.size()
+
+    query_states = self.q_proj(hidden_states)
+    key_states = self.k_proj(hidden_states)
+    value_states = self.v_proj(hidden_states)
+
+    query_states = query_states.view(bsz, q_len,
+                                     self.num_heads, self.head_dim).transpose(1, 2)
+    key_states = key_states.view(bsz, q_len,
+                                 self.num_key_value_heads, self.head_dim).transpose(1, 2)
+    value_states = value_states.view(bsz, q_len,
+                                     self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is not None:
+        invalidInputError(self.layer_idx is not None,
+                          "The cache structure has changed since version v4.36. "
+                          f"If you are using {self.__class__.__name__} "
+                          "for auto-regressive decoding with k/v caching, "
+                          "please make sure to initialize the attention class "
+                          "with a layer index.")
+        kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+
+    if use_fuse_rope:
+        query_states, key_states = apply_rotary_pos_emb_cache_freq_xpu(query_states, key_states,
+                                                                       sin, cos, "qwen2_moe",
+                                                                       position_ids)
+    else:
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states,
+                                                        cos, sin, position_ids)
+
+    if past_key_value is not None:
+        cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+        key_states, value_states = past_key_value.update(key_states, value_states,
+                                                         self.layer_idx, cache_kwargs)
+    if q_len == 1 and query_states.device.type == 'xpu' and not self.training \
+            and not hidden_states.requires_grad:
+        import linear_q4_0
+        attn_weights = linear_q4_0.query_key_fp8_matmul(query_states, key_states)
+    else:
+        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))
+
+    attn_weights = attn_weights / math.sqrt(self.head_dim)
+
+    invalidInputError(attn_weights.size() == (bsz, self.num_heads, q_len, kv_seq_len),
+                      ("Attention weights should be of size "
+                       f"{(bsz, self.num_heads, q_len, kv_seq_len)},"
+                       "but is {attn_weights.size()}"))
+
+    if attention_mask is not None:
+        invalidInputError(attention_mask.size() == (bsz, 1, q_len, kv_seq_len),
+                          (f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}"
+                           f" but is {attention_mask.size()}"))
+
+        attn_weights = attn_weights + attention_mask
+
+    # upcast attention to fp32
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1,
+                                         dtype=torch.float32).to(query_states.dtype)
+    attn_weights = nn.functional.dropout(attn_weights,
+                                         p=self.attention_dropout, training=self.training)
+    if q_len == 1 and query_states.device.type == 'xpu' and not self.training \
+            and not hidden_states.requires_grad:
+        import linear_q4_0
+        attn_output = linear_q4_0.attn_value_fp8_matmul(attn_weights,
+                                                        value_states.transpose(-1, -2))
+    else:
+        attn_output = torch.matmul(attn_weights, value_states)
+
+    invalidInputError(attn_output.size() == (bsz, self.num_heads, q_len, self.head_dim),
+                      "`attn_output` should be of size "
+                      f"{(bsz, self.num_heads, q_len, self.head_dim)},"
+                      f" but is {attn_output.size()}")
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+    attn_output = self.o_proj(attn_output)
+
+    if not output_attentions:
+        attn_weights = None
+
+    return attn_output, attn_weights, past_key_value
+
+
+def qwen2moe_attention_forward_origin(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+    **kwargs,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids)
+
+    if "padding_mask" in kwargs:
+        warnings.warn(
+            "Passing `padding_mask` is deprecated and will be removed in v4.37. "
+            "Please make sure use `attention_mask` instead.`"
+        )
+    bsz, q_len, _ = hidden_states.size()
+    device = hidden_states.device
+
+    enough_kv_room = is_enough_kv_cache_room_4_36(past_key_value, self.layer_idx)
+
+    qtype_check = decoding_fast_path_qtype_check(self.q_proj)
+    decoding_fast_path = (qtype_check and use_fuse_rope
+                          and enough_kv_room and bsz * q_len == 1)
+    decoding_fast_path = decoding_fast_path and not self.q_proj.enable_xetla
+    if decoding_fast_path:
+        hidden_states = hidden_states.view(1, -1)
+        cache_k = past_key_value.key_cache[self.layer_idx]
+        cache_v = past_key_value.value_cache[self.layer_idx]
+        kv_seq_len = cache_k.shape[-2]
+        import linear_q4_0
+        args = [hidden_states, self.q_proj.weight, self.k_proj.weight, self.v_proj.weight,
+                self.q_proj.bias, self.k_proj.bias, self.v_proj.bias, position_ids, cache_k,
+                cache_v, self.q_proj.weight.qtype, self.v_proj.weight.qtype, kv_seq_len,
+                self.head_dim, self.rotary_emb.base]
+        query_states, key_states, value_states = linear_q4_0.forward_qkv_bias(*args)
+        kv_seq_len += 1
+        if self.layer_idx == 0:
+            past_key_value._seen_tokens = kv_seq_len
+        past_key_value.key_cache[self.layer_idx] = key_states
+        past_key_value.value_cache[self.layer_idx] = value_states
+    else:
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len,
+                                         self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len,
+                                     self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len,
+                                         self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                invalidInputError(
+                    False,
+                    "The cache structure has changed since version v4.36. "
+                    f"If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, "
+                    "please make sure to initialize the attention class with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        if use_fuse_rope:
+            query_states, key_states = apply_rotary_pos_emb_cache_freq_xpu(query_states, key_states,
+                                                                           sin, cos, "qwen2_moe",
+                                                                           position_ids)
+        else:
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states,
+                                                            cos, sin, position_ids)
+        if past_key_value is not None:
+            if self.layer_idx == 0:
+                past_key_value._seen_tokens += key_states.shape[-2]
+
+            if len(past_key_value.key_cache) <= self.layer_idx:
+                past_key_value.key_cache.append(key_states)
+                past_key_value.value_cache.append(value_states)
+            else:
+                cache_k = past_key_value.key_cache[self.layer_idx]
+                cache_v = past_key_value.value_cache[self.layer_idx]
+
+                if not enough_kv_room:
+                    # allocate new
+                    new_c_k, new_c_v = extend_kv_cache(bsz,
+                                                       self.num_key_value_heads,  # Support GQA
+                                                       self.head_dim,
+                                                       cache_k.size(2),
+                                                       kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH,
+                                                       dtype=cache_k.dtype,
+                                                       device=device)
+
+                    new_c_k[:] = cache_k
+                    new_c_v[:] = cache_v
+                    cache_k = new_c_k
+                    cache_v = new_c_v
+
+                key_states, value_states = append_kv_cache(cache_k,
+                                                           cache_v,
+                                                           key_states,
+                                                           value_states)
+
+                # update past_key_value
+                past_key_value.key_cache[self.layer_idx] = key_states
+                past_key_value.value_cache[self.layer_idx] = value_states
+    # 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)
+
+    if not self.training and not hidden_states.requires_grad and \
+            use_flash_attention(query_states, key_states, attention_mask):
+        attn_output = F.scaled_dot_product_attention(query_states.to(device, dtype=torch.float16),
+                                                     key_states.to(device, dtype=torch.float16),
+                                                     value_states.to(device, dtype=torch.float16),
+                                                     is_causal=True)
+        attn_weights = None
+    else:
+        attn_weights = torch.matmul(query_states,
+                                    key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        invalidInputError(attn_weights.size() == (bsz, self.num_heads, q_len, kv_seq_len),
+                          ("Attention weights should be of size "
+                           f"{(bsz, self.num_heads, q_len, kv_seq_len)},"
+                           "but is {attn_weights.size()}"))
+
+        if attention_mask is not None:
+            invalidInputError(attention_mask.size() == (bsz, 1, q_len, kv_seq_len),
+                              (f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}"
+                               f" but is {attention_mask.size()}"))
+
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights,
+                                             dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights,
+                                             p=self.attention_dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+    invalidInputError(attn_output.size() == (bsz, self.num_heads, q_len, self.head_dim),
+                      "`attn_output` should be of size "
+                      f"{(bsz, self.num_heads, q_len, self.head_dim)},"
+                      f" but is {attn_output.size()}")
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+    attn_output = self.o_proj(attn_output)
+
+    if not output_attentions:
+        attn_weights = None
+
+    return attn_output.to(hidden_states.dtype), attn_weights, past_key_value
+
+
+def qwen2moe_attention_forward_sdpa(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+    **kwargs,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids)
+
+    if "padding_mask" in kwargs:
+        warnings.warn(
+            "Passing `padding_mask` is deprecated and will be removed in v4.37. "
+            "Please make sure use `attention_mask` instead.`"
+        )
+    bsz, q_len, _ = hidden_states.size()
+    device = hidden_states.device
+
+    enough_kv_room = is_enough_kv_cache_room_4_36(past_key_value, self.layer_idx)
+
+    qtype_check = decoding_fast_path_qtype_check(self.q_proj)
+    decoding_fast_path = (qtype_check and use_fuse_rope
+                          and enough_kv_room and bsz * q_len == 1)
+    decoding_fast_path = decoding_fast_path and not self.q_proj.enable_xetla
+    if decoding_fast_path:
+        hidden_states = hidden_states.view(1, -1)
+        cache_k = past_key_value.key_cache[self.layer_idx]
+        cache_v = past_key_value.value_cache[self.layer_idx]
+        kv_seq_len = cache_k.shape[-2]
+        import linear_q4_0
+        args = [hidden_states, self.q_proj.weight, self.k_proj.weight, self.v_proj.weight,
+                self.q_proj.bias, self.k_proj.bias, self.v_proj.bias, position_ids, cache_k,
+                cache_v, self.q_proj.weight.qtype, self.v_proj.weight.qtype, kv_seq_len,
+                self.head_dim, self.rotary_emb.base]
+        query_states, key_states, value_states = linear_q4_0.forward_qkv_bias(*args)
+        kv_seq_len += 1
+        if self.layer_idx == 0:
+            past_key_value._seen_tokens = kv_seq_len
+        past_key_value.key_cache[self.layer_idx] = key_states
+        past_key_value.value_cache[self.layer_idx] = value_states
+    else:
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len,
+                                         self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len,
+                                     self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len,
+                                         self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                invalidInputError(
+                    False,
+                    "The cache structure has changed since version v4.36. "
+                    f"If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, "
+                    "please make sure to initialize the attention class with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        if use_fuse_rope:
+            query_states, key_states = apply_rotary_pos_emb_cache_freq_xpu(query_states, key_states,
+                                                                           sin, cos, "qwen2_moe",
+                                                                           position_ids)
+        else:
+            query_states, key_states = apply_rotary_pos_emb(query_states, key_states,
+                                                            cos, sin, position_ids)
+        if past_key_value is not None:
+            if self.layer_idx == 0:
+                past_key_value._seen_tokens += key_states.shape[-2]
+
+            if len(past_key_value.key_cache) <= self.layer_idx:
+                past_key_value.key_cache.append(key_states)
+                past_key_value.value_cache.append(value_states)
+            else:
+                cache_k = past_key_value.key_cache[self.layer_idx]
+                cache_v = past_key_value.value_cache[self.layer_idx]
+
+                if not enough_kv_room:
+                    # allocate new
+                    new_c_k, new_c_v = extend_kv_cache(bsz,
+                                                       self.num_key_value_heads,  # Support GQA
+                                                       self.head_dim,
+                                                       cache_k.size(2),
+                                                       kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH,
+                                                       dtype=cache_k.dtype,
+                                                       device=device)
+
+                    new_c_k[:] = cache_k
+                    new_c_v[:] = cache_v
+                    cache_k = new_c_k
+                    cache_v = new_c_v
+
+                key_states, value_states = append_kv_cache(cache_k,
+                                                           cache_v,
+                                                           key_states,
+                                                           value_states)
+
+                # update past_key_value
+                past_key_value.key_cache[self.layer_idx] = key_states
+                past_key_value.value_cache[self.layer_idx] = value_states
+    # 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)
+
+    if output_attentions:
+        attn_weights = torch.matmul(query_states,
+                                    key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        invalidInputError(attn_weights.size() == (bsz, self.num_heads, q_len, kv_seq_len),
+                          ("Attention weights should be of size "
+                           f"{(bsz, self.num_heads, q_len, kv_seq_len)},"
+                           "but is {attn_weights.size()}"))
+
+        if attention_mask is not None:
+            invalidInputError(attention_mask.size() == (bsz, 1, q_len, kv_seq_len),
+                              (f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}"
+                               f" but is {attention_mask.size()}"))
+
+            attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights,
+                                             dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights,
+                                             p=self.attention_dropout, training=self.training)
+    else:
+        attn_weights = None
+
+    from torch.nn.functional import scaled_dot_product_attention as sdpa
+    attn_output = sdpa(query_states,
+                       key_states,
+                       value_states,
+                       attn_mask=attention_mask,
+                       dropout_p=self.attention_dropout if self.training else 0.0,
+                       is_causal=self.is_causal and attention_mask is None and q_len > 1)
+
+    invalidInputError(attn_output.size() == (bsz, self.num_heads, q_len, self.head_dim),
+                      "`attn_output` should be of size "
+                      f"{(bsz, self.num_heads, q_len, self.head_dim)},"
+                      f" but is {attn_output.size()}")
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+    attn_output = self.o_proj(attn_output)
+
+    return attn_output, attn_weights, past_key_value
+
+
+def qwen2moe_moeblock_forward(self, hidden_states: torch.Tensor):
+    batch_size, sequence_length, hidden_dim = hidden_states.shape
+    hidden_states = hidden_states.view(-1, hidden_dim)
+    bs = hidden_states.shape[0]
+    # router_logits: (batch * sequence_length, n_experts)
+    router_logits = self.gate(hidden_states)
+
+    routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+    routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+    if self.norm_topk_prob:
+        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+    # we cast back to the input dtype
+    routing_weights = routing_weights.to(hidden_states.dtype)
+
+    if bs == 1:
+        selected_experts = selected_experts[0].cpu().tolist()
+        for idx in range(self.top_k):
+            exp_id = selected_experts[idx]
+            expert_layer = self.experts[exp_id]
+            weight = routing_weights[:, idx]
+            if idx == 0:
+                final_hidden_states = expert_layer(hidden_states) * weight
+            else:
+                final_hidden_states = final_hidden_states + expert_layer(hidden_states) * weight
+    elif bs < 256:
+        final_hidden_states = torch.zeros((batch_size * sequence_length, hidden_dim),
+                                          dtype=hidden_states.dtype, device=hidden_states.device)
+        import linear_q4_0
+        indexes = linear_q4_0.get_moe_indexes(selected_experts.to(torch.int32).cpu(), 60)
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx_list = indexes[0][expert_idx]
+            top_x_list = indexes[1][expert_idx]
+            if len(idx_list) == 0:
+                continue
+
+            top_x = torch.tensor(top_x_list, dtype=torch.long, device=hidden_states.device)
+            current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * \
+                routing_weights[top_x_list, idx_list, None]
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+    else:
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim),
+            dtype=hidden_states.dtype,
+            device=hidden_states.device
+        )
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(selected_experts,
+                                                  num_classes=self.num_experts).permute(2, 1, 0)
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            if top_x.shape[0] == 0:
+                continue
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+    shared_expert_output = self.shared_expert(hidden_states)
+    shared_expert_output = F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_expert_output
+
+    final_hidden_states = final_hidden_states + shared_expert_output
+
+    final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+    return final_hidden_states, router_logits

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

@@ -169,7 +169,7 @@ def rotate_every_two(x):
 
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, model_family):
     if model_family in ["llama", "baichuan", "internlm", "aquila", "gpt_neox", "mistral",
-                        "mixtral", "qwen2", "yuan", "stablelm"]:
+                        "mixtral", "qwen2", "yuan", "stablelm", "qwen2_moe"]:
         # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
         cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]
         sin = sin.squeeze(1).squeeze(0)  # [seq_len, dim]
@@ -226,7 +226,7 @@ def apply_rotary_pos_emb_cache_freq_xpu(q, k, sin, cos, model_family, position_i
     k_embed = torch.empty(k.shape, dtype=k.dtype, device=k.device)
     if model_family in ["qwen", "mixtral"]:
         linear_q4_0.apply_rotary_embedding_half_q_and_k_cache_freq(q, k, sin, cos, q_embed, k_embed)
-    elif model_family in ["qwen2", "yuan", "stablelm"]:
+    elif model_family in ["qwen2", "yuan", "stablelm", "qwen2_moe"]:
         cos = cos.to(q.dtype)
         sin = sin.to(q.dtype)
         cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]