refactor to simplify following upgrade 2 (#12685)

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

@@ -1590,6 +1590,9 @@ def _optimize_post(model):
         convert_forward(model,
                         module.Qwen2ForCausalLM,
                         qwen2_causal_lm_forward)
+        convert_forward(model,
+                        module.Qwen2Model,
+                        qwen2_model_forward)
         convert_forward(model,
                         module.Qwen2RMSNorm,
                         rms_norm_forward)
@@ -1602,12 +1605,6 @@ def _optimize_post(model):
         convert_forward(model,
                         module.Qwen2SdpaAttention,
                         qwen2_attention_forward)
-        if version.parse(trans_version) >= version.parse("4.42"):
-            from ipex_llm.transformers.models.qwen2 import qwen2_model_forward_4_42
-            convert_forward(model, module.Qwen2Model, qwen2_model_forward_4_42)
-        else:
-            from ipex_llm.transformers.models.qwen2 import qwen2_model_forward
-            convert_forward(model, module.Qwen2Model, qwen2_model_forward)
     elif model.config.model_type == "qwen2_moe":
         # for Qwen1.5-MOE-A2.7B
         modeling_module_name = model.__class__.__module__
@@ -1819,9 +1816,7 @@ def _optimize_post(model):
         from ipex_llm.transformers.models.phi3 import attention_forward
         convert_forward(model, module.Phi3Attention, attention_forward)
         convert_forward(model, module.Phi3SdpaAttention, attention_forward)
-        from ipex_llm.transformers.models.phi3 import mlp_forward
+        convert_forward(model, module.Phi3MLP, mlp_silu_forward)
-        convert_forward(model, module.Phi3MLP, mlp_forward)
-        from ipex_llm.transformers.models.common import rms_norm_forward
         convert_forward(model, module.Phi3RMSNorm, rms_norm_forward)
         if model.config.model_type == "phi3":
             from ipex_llm.transformers.models.phi3 import phi3_model_forward_wrapper

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

@@ -30,8 +30,7 @@ from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp
 from ipex_llm.transformers.models.utils import update_past_key_value
 from ipex_llm.transformers.models.utils import should_use_fuse_rope
 from ipex_llm.transformers.models.utils import use_sdp
-from ipex_llm.transformers.models.utils import apply_rotary_pos_emb, SILU
+from ipex_llm.transformers.models.utils import apply_rotary_pos_emb
-from ipex_llm.transformers.models.utils import mlp_fusion_check
 from ipex_llm.transformers.models.utils import is_enough_kv_cache_room_4_36
 from ipex_llm.transformers.kv import DynamicCompressFp8Cache, DynamicCompressCache
 import warnings

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

@@ -113,21 +113,6 @@ def internlm_attention_forward(
     return attn_output, attn_weights, past_key_value
 
 
-def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
-    """
-    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep).
-    The hidden states go from (batch,
-    num_key_value_heads, seqlen, head_dim) to
-    (batch, num_attention_heads, seqlen, head_dim)
-    """
-    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
-    if n_rep == 1:
-        return hidden_states
-    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads,
-                                                           n_rep, slen, head_dim)
-    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
-
-
 def internlm2_attention_forward(
     self,
     hidden_states: torch.Tensor,

python/llm/src/ipex_llm/transformers/models/phi3.py

@@ -39,7 +39,6 @@ import warnings
 from ipex_llm.transformers.models.common import attention_softmax
 from ipex_llm.transformers.models.common import scaled_dot_product_attention
 from ipex_llm.transformers.models.utils import should_use_fuse_rope, rotate_half
-from ipex_llm.transformers.models.utils import mlp_fusion_check, SILU
 from ipex_llm.transformers.models.utils import use_sdp, use_sdp_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_compresskv, is_enough_kv_cache_room_4_36
@@ -213,24 +212,8 @@ def split_mlp(module: torch.nn.Module):
 
         del module.gate_up_proj
 
-
+        # rename activation function
-def mlp_forward(
+        module.act_fn = module.activation_fn
-    self,
-    hidden_states: torch.FloatTensor
-) -> torch.FloatTensor:
-    x_2d = hidden_states.view(-1, hidden_states.shape[-1])
-    qtype = getattr(self.gate_proj, "qtype", None)
-    if mlp_fusion_check(x_2d, qtype, self.training):
-        x_2d = x_2d.contiguous()
-        import xe_linear
-        return self.down_proj(xe_linear.mlp_forward_xpu(
-            x_2d, self.gate_proj.weight.data, self.up_proj.weight.data,
-            x_2d.shape[0], x_2d.shape[1], self.gate_proj.out_features,
-            SILU, qtype
-        ))
-    return self.down_proj(
-        self.activation_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states)
-    )
 
 
 def phi3_model_forward_wrapper(origin_model_forward):

python/llm/src/ipex_llm/transformers/models/qwen2.py

@@ -51,16 +51,11 @@ from ipex_llm.transformers.models.utils import use_quantize_kv_cache, \
     should_use_compresskv, is_enough_kv_cache_room_4_36
 from ipex_llm.transformers.kv import DynamicFp8Cache, DynamicNormalCache, \
     DynamicCompressCache, DynamicCompressFp8Cache
-from ipex_llm.utils.common import invalidInputError
 
-from transformers.models.qwen2.modeling_qwen2 import Qwen2Attention, Qwen2MLP
+from transformers.models.qwen2.modeling_qwen2 import Qwen2Model, Qwen2Attention, Qwen2MLP
 from transformers.models.qwen2.modeling_qwen2 import apply_rotary_pos_emb
 from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from transformers.cache_utils import Cache
-from transformers import logging
-
-
-logger = logging.get_logger(__name__)
 
 
 def qwen2_model_forward(
@@ -74,50 +69,18 @@ def qwen2_model_forward(
     output_attentions: Optional[bool] = None,
     output_hidden_states: Optional[bool] = None,
     return_dict: Optional[bool] = None,
-    cache_position: Optional[torch.LongTensor] = None,  # for transformers >= 4.42
+    cache_position: Optional[torch.LongTensor] = None,
 ) -> Union[Tuple, BaseModelOutputWithPast]:
-    output_attentions = (
+    # IPEX-LLM OPT start: kv cache and quantize kv cache
-        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
-
-    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-
-    # retrieve input_ids and inputs_embeds
-    if input_ids is not None and inputs_embeds is not None:
-        invalidInputError(False,
-                          "You cannot specify both input_ids and inputs_embeds at the same time")
-    elif input_ids is not None:
-        batch_size, seq_length = input_ids.shape
-    elif inputs_embeds is not None:
-        batch_size, seq_length, _ = inputs_embeds.shape
-    else:
-        invalidInputError(False,
-                          "You have to specify either decoder_input_ids or decoder_inputs_embeds")
-
-    if self.gradient_checkpointing and self.training:
-        if use_cache:
-            logger.warning_once(
-                "`use_cache=True` is incompatible with gradient checkpointing. "
-                "Setting `use_cache=False`..."
-            )
-            use_cache = False
-
-    past_key_values_length = 0
-
-    # ipex-llm changes start
-    # IPEX-LLM OPT: kv cache and quantize kv cache
     inputs = input_ids if input_ids is not None else inputs_embeds
-    num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
-    use_quantize_kv = (
+    use_cache = True if inputs.device.type == "xpu" else use_cache
-        self.config.hidden_size != 3584     # disable quantize kv in specific model
+
-        and use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs, num_heads, num_kv_heads)
+    use_quantize_kv = self.config.hidden_size != 3584 and use_quantize_kv_cache(
+        self.layers[0].mlp.down_proj, inputs,
+        self.config.num_attention_heads, self.config.num_key_value_heads
     )
+
     use_compress_kv = should_use_compresskv(inputs, inputs.shape[1]) or \
         isinstance(past_key_values, DynamicCompressCache)
 
@@ -133,274 +96,26 @@ def qwen2_model_forward(
         if not use_quantize_kv and not use_compress_kv and not isinstance(past_key_values,
                                                                           DynamicNormalCache):
             past_key_values = DynamicNormalCache.from_legacy_cache(past_key_values)
-        past_key_values_length = past_key_values.get_usable_length(seq_length)
     # ipex-llm changes end
 
-    if position_ids is None:
+    # `cache_position` is required after transformers 4.42
-        device = input_ids.device if input_ids is not None else inputs_embeds.device
+    if cache_position is not None:
-        position_ids = torch.arange(
+        kwargs = {"cache_position": cache_position}
-            past_key_values_length, seq_length + past_key_values_length,
-            dtype=torch.long, device=device
-        )
-        position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
     else:
-        position_ids = position_ids.view(-1, seq_length).long()
+        kwargs = {}
 
-    if inputs_embeds is None:
+    return Qwen2Model.forward(
-        inputs_embeds = self.embed_tokens(input_ids)
+        self=self,
-
+        input_ids=input_ids,
-    flash_attn_2 = self._attn_implementation == "flash_attention_2"
-    if attention_mask is not None and flash_attn_2 and use_cache:
-
-        is_padding_right = attention_mask[:, -1].sum().item() != batch_size
-        if is_padding_right:
-            invalidInputError(
-                False,
-                "You are attempting to perform batched generation with padding_side='right'"
-                " this may lead to unexpected behaviour for Flash Attention version of Qwen2."
-                " Make sure to  call `tokenizer.padding_side  = 'left'` before tokenizing "
-                "the input. "
-            )
-
-    from transformers.models.qwen2.modeling_qwen2 import _prepare_4d_causal_attention_mask_for_sdpa
-    from transformers.models.qwen2.modeling_qwen2 import _prepare_4d_causal_attention_mask
-
-    # ipex-llm changes start: don't generate `attention_mask` in decode phase
-    if seq_length == 1:
-        attention_mask = None
-    # ipex-llm changes end
-    elif self._attn_implementation == "flash_attention_2":
-        # 2d mask is passed through the layers
-        attention_mask = attention_mask if (attention_mask is not None and
-                                            0 in attention_mask) else None
-    elif self._attn_implementation == "sdpa" and not output_attentions:
-        # output_attentions=True can not be supported when using SDPA, and we fall back on
-        # the manual implementation that requires a 4D causal mask in all cases.
-        attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
-            attention_mask,
-            (batch_size, seq_length),
-            inputs_embeds,
-            past_key_values_length,
-        )
-    else:
-        # 4d mask is passed through the layers
-        attention_mask = _prepare_4d_causal_attention_mask(
-            attention_mask,
-            (batch_size, seq_length),
-            inputs_embeds,
-            past_key_values_length,
-            sliding_window=self.config.sliding_window,
-        )
-
-    hidden_states = inputs_embeds
-
-    # decoder layers
-    all_hidden_states = () if output_hidden_states else None
-    all_self_attns = () if output_attentions else None
-    next_decoder_cache = None
-
-    for decoder_layer in self.layers:
-        if output_hidden_states:
-            all_hidden_states += (hidden_states,)
-
-        if self.gradient_checkpointing and self.training:
-            layer_outputs = self._gradient_checkpointing_func(
-                decoder_layer.__call__,
-                hidden_states,
-                attention_mask,
-                position_ids,
-                past_key_values,
-                output_attentions,
-                use_cache,
-            )
-        else:
-            # ipex-llm changes
-            curr_device = decoder_layer.input_layernorm.weight.device
-            if attention_mask is not None:
-                attention_mask = attention_mask.to(curr_device)
-            if position_ids is not None:
-                position_ids = position_ids.to(curr_device)
-            # ipex-llm changes end
-            layer_outputs = decoder_layer(
-                hidden_states,
         attention_mask=attention_mask,
         position_ids=position_ids,
-                past_key_value=past_key_values,
+        past_key_values=past_key_values,
-                output_attentions=output_attentions,
+        inputs_embeds=inputs_embeds,
         use_cache=use_cache,
-            )
-
-        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,)
-
-    # ipex-llm changes start: remove `to_legacy_cache`
-    next_cache = None
-    if use_cache:
-        next_cache = next_decoder_cache
-    # ipex-llm changes end
-
-    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 qwen2_model_forward_4_42(
-    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]:
-    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
-
-    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 self.gradient_checkpointing and self.training:
-        if use_cache:
-            logger.warning_once(
-                "`use_cache=True` is incompatible with gradient checkpointing. "
-                "Setting `use_cache=False`..."
-            )
-            use_cache = False
-
-    if inputs_embeds is None:
-        inputs_embeds = self.embed_tokens(input_ids)
-
-    # ipex-llm changes start
-    # IPEX-LLM OPT: kv cache and quantize kv cache
-    num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
-    use_quantize_kv = (
-        self.config.hidden_size != 3584     # disable quantize kv in specific model
-        and use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs_embeds,
-                                  num_heads, num_kv_heads)
-    )
-    use_compress_kv = should_use_compresskv(inputs_embeds, inputs_embeds.shape[1]) or \
-        isinstance(past_key_values, DynamicCompressCache)
-
-    if use_cache:
-        if use_compress_kv and not isinstance(past_key_values, DynamicCompressCache):
-            if use_quantize_kv:
-                past_key_values = DynamicCompressFp8Cache.from_legacy_cache(past_key_values)
-            else:
-                past_key_values = DynamicCompressCache.from_legacy_cache(past_key_values)
-        elif use_quantize_kv and not use_compress_kv and not isinstance(past_key_values,
-                                                                        DynamicFp8Cache):
-            past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)
-        if not use_quantize_kv and not use_compress_kv and not isinstance(past_key_values,
-                                                                          DynamicNormalCache):
-            past_key_values = DynamicNormalCache.from_legacy_cache(past_key_values)
-    # ipex-llm changes end
-
-    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
-    )
-
-    hidden_states = inputs_embeds
-
-    # decoder layers
-    all_hidden_states = () if output_hidden_states else None
-    all_self_attns = () if output_attentions else None
-    next_decoder_cache = None
-
-    for decoder_layer in self.layers:
-        if output_hidden_states:
-            all_hidden_states += (hidden_states,)
-
-        if self.gradient_checkpointing and self.training:
-            layer_outputs = self._gradient_checkpointing_func(
-                decoder_layer.__call__,
-                hidden_states,
-                causal_mask,
-                position_ids,
-                past_key_values,
-                output_attentions,
-                use_cache,
-                cache_position,
-            )
-        else:
-            layer_outputs = decoder_layer(
-                hidden_states,
-                attention_mask=causal_mask,
-                position_ids=position_ids,
-                past_key_value=past_key_values,
         output_attentions=output_attentions,
-                use_cache=use_cache,
+        output_hidden_states=output_hidden_states,
-                cache_position=cache_position,
+        return_dict=return_dict,
-            )
+        **kwargs
-
-        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,)
-
-    # ipex-llm changes start: remove `to_legacy_cache`
-    next_cache = None
-    if use_cache:
-        next_cache = next_decoder_cache
-    # ipex-llm changes end
-
-    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,
     )
 
 

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

@@ -272,26 +272,6 @@ def use_xmx(x: torch.Tensor, qtype: int):
     )
 
 
-def fp16_fusion_check(proj, x, training):
-    # only use fp16 fusion on PVC inference
-    if proj is None:
-        return False
-    if not hasattr(proj, "qtype"):
-        return False
-    if proj.qtype != ggml_tensor_qtype["fp16"]:
-        return False
-    if proj.weight_type != 2:
-        return False
-    if training:
-        return False
-    if x.requires_grad:
-        return False
-    device_type = get_xpu_device_name(x.device)
-    if device_type != "pvc":
-        return False
-    return True
-
-
 def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     batch, num_key_value_heads, slen, head_dim = hidden_states.shape
     if n_rep == 1: