refactor yuan2 (#11235)

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

@@ -682,39 +682,8 @@ def _optimize_pre(model):
                 model.lm_head.weight.data = norm_weight
     # for yuan 2.0
     if model.config.model_type == "yuan":
-        def merge_qk_proj_func(module):
+        from ipex_llm.transformers.models.yuan import merge_qk
-            if "YuanAttention" in module.__class__.__name__:
+        model.apply(merge_qk)
-                q_weight = module.q_proj.weight.data
-                k_weight = module.k_proj.weight.data
-                num_heads = module.num_heads
-                head_dim = module.head_dim
-                hidden_size = module.hidden_size
-
-                weight_q = torch.cat([
-                    q_weight.view(num_heads, head_dim, hidden_size)[0::2, :, :],
-                    k_weight.view(num_heads, head_dim, hidden_size)[0::2, :, :],
-                ], dim=0).view(num_heads * head_dim, hidden_size)
-
-                weight_k = torch.cat([
-                    q_weight.view(num_heads, head_dim, hidden_size)[1::2, :, :],
-                    k_weight.view(num_heads, head_dim, hidden_size)[1::2, :, :],
-                ], dim=0).view(num_heads * head_dim, hidden_size)
-
-                merged_q_proj = torch.nn.Linear(0, 0, False)
-                merged_q_proj.weight = torch.nn.Parameter(weight_q, requires_grad=False)
-                merged_q_proj.in_features = hidden_size
-                merged_q_proj.out_features = num_heads * head_dim
-                module.merged_q_proj = merged_q_proj
-
-                merged_k_proj = torch.nn.Linear(0, 0, False)
-                merged_k_proj.weight = torch.nn.Parameter(weight_k, requires_grad=False)
-                merged_k_proj.in_features = hidden_size
-                merged_k_proj.out_features = num_heads * head_dim
-                module.merged_k_proj = merged_k_proj
-
-                del module.q_proj
-                del module.k_proj
-        model.apply(merge_qk_proj_func)
     # for bge-large
     if model.config.model_type == 'bert' and (
         not model.config.is_decoder and

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

@@ -20,32 +20,41 @@
 # https://huggingface.co/IEITYuan/Yuan2-2B-hf/blob/7ab7b3c18eb8e5232ce2a3f720d4e6f4b53a2806/README.md#%E5%A3%B0%E6%98%8E%E4%B8%8E%E5%8D%8F%E8%AE%AEterms-and-conditions
 #
 
-import copy
 import math
-from einops import rearrange
 from typing import Optional, Tuple
 
 import torch
-import torch.nn as nn
 
 from ipex_llm.utils.common import invalidInputError
 from ipex_llm.transformers.models.utils import apply_rotary_pos_emb, \
-    apply_rotary_pos_emb_cache_freq_xpu, mlp_fusion_check, fp16_fusion_check
+    mlp_fusion_check, fp16_fusion_check
-from ipex_llm.transformers.models.utils import init_kv_cache, extend_kv_cache, append_kv_cache
+from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp8_kv_cache
-from ipex_llm.transformers.models.utils import init_fp8_kv_cache, append_fp8_kv_cache, \
+from ipex_llm.transformers.models.utils import SILU, update_past_key_value
-    restore_fp8_kv_cache, use_quantize_kv_cache
+from ipex_llm.transformers.models.utils import should_use_fuse_rope, use_sdp, use_sdp_causal
-from ipex_llm.transformers.models.utils import is_enough_kv_cache_room_4_31, SILU
-
-import os
-
-KV_CACHE_ALLOC_BLOCK_LENGTH = int(os.environ.get("KV_CACHE_ALLOC_BLOCK_LENGTH", 256))
 
 
-def should_use_fuse_rope(self, hidden_states, position_ids):
+def merge_qk(module: torch.nn.Module):
-    use_fuse_rope = hidden_states.device.type == "xpu"
+    if "YuanAttention" in module.__class__.__name__:
-    use_fuse_rope = use_fuse_rope and not (self.training and hidden_states.requires_grad)
+        q_weight = module.q_proj.weight.data
-    use_fuse_rope = use_fuse_rope and position_ids is not None
+        k_weight = module.k_proj.weight.data
-    return use_fuse_rope
+        num_heads = module.num_heads
+        head_dim = module.head_dim
+        hidden_size = module.hidden_size
+
+        merged_qk_proj = torch.nn.Linear(0, 0, False)
+        weight = torch.cat([
+            q_weight.view(num_heads, head_dim, hidden_size)[0::2, :, :],
+            k_weight.view(num_heads, head_dim, hidden_size)[0::2, :, :],
+            q_weight.view(num_heads, head_dim, hidden_size)[1::2, :, :],
+            k_weight.view(num_heads, head_dim, hidden_size)[1::2, :, :],
+        ], dim=0).view(num_heads * head_dim * 2, hidden_size)
+        merged_qk_proj.weight = torch.nn.Parameter(weight, requires_grad=False)
+        merged_qk_proj.in_features = hidden_size
+        merged_qk_proj.out_features = num_heads * head_dim * 2
+        module.qk_proj = merged_qk_proj
+
+        del module.q_proj
+        del module.k_proj
 
 
 def yuan_localized_filtering_forward(
@@ -142,43 +151,14 @@ def yuan_attention_forward(
     past_key_value: Optional[Tuple[torch.Tensor]] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-    if use_quantize_kv_cache(self.merged_q_proj, hidden_states):
-        forward_function = yuan_attention_forward_quantized
-    else:
-        forward_function = yuan_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,
-    )
-
-
-def yuan_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,
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
     bsz, q_len, _ = hidden_states.size()
     device = hidden_states.device
-    before_hidden_states = None
-    is_first_step = False
-
-    use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids)
 
     invalidInputError(use_cache, "use_cache=True is needed")
     invalidInputError(not self.use_shareqk, "use_shareqk is not supported for now")
 
     if past_key_value is None:
-        is_first_step = True
         if q_len >= 2:
             before_hidden_states = hidden_states[:, -2:, :].transpose(0, 1).half()
         else:
@@ -193,112 +173,75 @@ def yuan_attention_forward_quantized(
         ], dim=0)
         before_hidden_states = this_hidden_states[-2:, :, ]
 
-    value_states = \
+    value_states = self.v_proj(hidden_states)
-        self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+    value_states = value_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
 
-    if is_first_step:
+    if past_key_value is None:
         hidden_states = yuan_localized_filtering_forward(self.lf_gate, hidden_states,
                                                          None, hidden_states.dtype)
     else:
         hidden_states = yuan_localized_filtering_forward(self.lf_gate, hidden_states,
                                                          this_hidden_states, hidden_states.dtype)
-    query_states = self.merged_q_proj(hidden_states)
+
-    key_states = self.merged_k_proj(hidden_states)
+    qk_states = self.qk_proj(hidden_states)
-    query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+    qk_states = qk_states.view(bsz, q_len, self.num_heads * 2, self.head_dim)
-    key_states = key_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+    qk_states = qk_states.transpose(1, 2)
+    query_states, key_states = torch.chunk(qk_states, 2, dim=1)
 
     kv_seq_len = key_states.shape[-2]
     if past_key_value is not None:
         kv_seq_len += past_key_value[0].shape[-2]
 
-    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    if should_use_fuse_rope(hidden_states, position_ids, self.training):
-    if use_fuse_rope:
+        import xe_addons
-        query_states, key_states = apply_rotary_pos_emb_cache_freq_xpu(query_states,
+        xe_addons.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids,
-                                                                       key_states,
+                                       query_states, key_states)
-                                                                       sin, cos,
-                                                                       "yuan",
-                                                                       position_ids)
     else:
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
         query_states, key_states = apply_rotary_pos_emb(query_states,
                                                         key_states,
                                                         cos, sin,
                                                         position_ids,
                                                         "yuan")
 
-    if past_key_value is None:
+    # IPEX-LLM OPT: kv cache and quantzie kv cache
-        # should use origin attn here
+    use_quantize_kv = use_quantize_kv_cache(self.qk_proj, hidden_states)
-        attn_weights = torch.matmul(query_states,
+    key_states, value_states = update_past_key_value(
-                                    key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+        None if past_key_value is None else (past_key_value[0], past_key_value[1]),
-
+        key_states, value_states,
-        invalidInputError(attn_weights.size() == (bsz, self.num_heads, q_len, kv_seq_len),
+        kv_seq_len, use_quantize_kv, device
-                          "Attention weights should be of size "
+    )
-                          f"{(bsz, self.num_heads, q_len, kv_seq_len)}, "
+    past_key_value = (key_states, value_states, before_hidden_states) if use_cache else None
-                          f"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
-            attn_weights = torch.max(attn_weights,
-                                     torch.tensor(torch.finfo(attn_weights.dtype).min))
-
-        # upcast attention to fp32
-        attn_weights = nn.functional.softmax(attn_weights, dim=-1,
-                                             dtype=torch.float32).to(query_states.dtype)
-        attn_output = torch.matmul(attn_weights, value_states)
-
-        if use_cache:
-            k_cache, v_cache = init_fp8_kv_cache(
-                bsz, self.num_heads, kv_seq_len, self.head_dim, device=device
-            )
-            key_states, value_states = append_fp8_kv_cache(k_cache, v_cache,
-                                                           key_states, value_states)
-            past_key_value = (key_states, value_states, before_hidden_states)
 
+    # IPEX-LLM OPT: sdp
+    if use_sdp(q_len, kv_seq_len, self.head_dim, query_states):
+        import xe_addons
+        if use_quantize_kv:
+            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 use_quantize_kv:
+            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:
-        k_cache, v_cache, _ = past_key_value
+        if use_quantize_kv:
-        key_states, value_states = append_fp8_kv_cache(k_cache, v_cache,
-                                                       key_states, value_states)
-        past_key_value = (key_states, value_states, before_hidden_states)
-
-        # torch.matmul
-        if query_states.size(2) != 1 or device.type != 'xpu':
             key_states, value_states = restore_fp8_kv_cache(key_states, value_states,
                                                             query_states.dtype)
-            attn_weights = torch.matmul(query_states, key_states.transpose(2, 3))
+        attn_weights = torch.matmul(query_states,
-        else:
+                                    key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
-            import xe_addons
-            attn_weights = xe_addons.query_key_fp8_matmul(query_states, key_states)
-
-        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)}, "
-                          f"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
-            attn_weights = torch.max(attn_weights,
-                                     torch.tensor(torch.finfo(attn_weights.dtype).min))
-
         # upcast attention to fp32
-        attn_weights = nn.functional.softmax(attn_weights, dim=-1,
+        attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1,
-                                             dtype=torch.float32).to(query_states.dtype)
+                                                   dtype=torch.float32).to(value_states.dtype)
-        if query_states.size(2) != 1 or device.type != 'xpu':
+        attn_output = torch.matmul(attn_weights, value_states)
-            attn_output = torch.matmul(attn_weights, value_states)
-        else:
-            import xe_addons
-            attn_output = xe_addons.attn_value_fp8_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)
     attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
@@ -307,146 +250,3 @@ def yuan_attention_forward_quantized(
         attn_weights = None
 
     return attn_output, attn_weights, past_key_value
-
-
-def yuan_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,
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-    use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids)
-    bsz, q_len, _ = hidden_states.size()
-    device = hidden_states.device
-    before_hidden_states = None
-    is_first_step = False
-    self.use_shareqk = False
-
-    enough_kv_room = is_enough_kv_cache_room_4_31(past_key_value)
-
-    invalidInputError(use_cache, "use_cache=True is needed")
-    invalidInputError(not self.use_shareqk, "use_shareqk is not supported for now")
-
-    if past_key_value is None:
-        is_first_step = True
-        if q_len >= 2:
-            before_hidden_states = hidden_states[:, -2:, :].transpose(0, 1).half()
-        else:
-            before_hidden_states = torch.zeros(2, bsz, self.hidden_size,
-                                               dtype=torch.half, device=hidden_states.device)
-            before_hidden_states[-1:, :, :] = hidden_states[:, -1:, :].transpose(0, 1)
-    else:
-        before_hidden_states = past_key_value[2]
-        this_hidden_states = torch.cat([
-            before_hidden_states,
-            hidden_states.transpose(0, 1).half(),
-        ], dim=0)
-        before_hidden_states = this_hidden_states[-2:, :, ]
-
-    value_states = \
-        self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-
-    if is_first_step:
-        hidden_states = yuan_localized_filtering_forward(self.lf_gate, hidden_states,
-                                                         None, hidden_states.dtype)
-    else:
-        hidden_states = yuan_localized_filtering_forward(self.lf_gate, hidden_states,
-                                                         this_hidden_states, hidden_states.dtype)
-    query_states = self.merged_q_proj(hidden_states)
-    key_states = self.merged_k_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_heads, self.head_dim).transpose(1, 2)
-
-    kv_seq_len = key_states.shape[-2]
-    if past_key_value is not None:
-        kv_seq_len += past_key_value[0].shape[-2]
-
-    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,
-                                                                       "yuan",
-                                                                       position_ids)
-    else:
-        query_states, key_states = apply_rotary_pos_emb(query_states,
-                                                        key_states,
-                                                        cos, sin,
-                                                        position_ids,
-                                                        "yuan")
-
-    if past_key_value is not None:
-        # reuse k, v, self_attention
-        cache_k = past_key_value[0]
-        cache_v = past_key_value[1]
-        if not enough_kv_room:
-            # allocate new
-            new_cache_k, new_cache_v = extend_kv_cache(bsz,
-                                                       self.num_heads,
-                                                       self.head_dim,
-                                                       cache_k.size(2),
-                                                       kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH,
-                                                       dtype=cache_k.dtype,
-                                                       device=device)
-            new_cache_k[:] = cache_k
-            new_cache_v[:] = cache_v
-            cache_k = new_cache_k
-            cache_v = new_cache_v
-
-        key_states, value_states = append_kv_cache(cache_k, cache_v, key_states, value_states)
-
-    elif use_cache:
-        max_cache_length = kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH
-        new_key_states, new_value_states = init_kv_cache(bsz,
-                                                         self.num_heads,
-                                                         self.head_dim,
-                                                         kv_seq_len,
-                                                         max_cache_length,
-                                                         dtype=key_states.dtype,
-                                                         device=device)
-        new_key_states[:] = key_states
-        new_value_states[:] = value_states
-        key_states = new_key_states
-        value_states = new_value_states
-
-    past_key_value = \
-        (key_states, value_states, before_hidden_states) if use_cache else None
-
-    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)}, "
-                      f"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
-        attn_weights = torch.max(attn_weights,
-                                 torch.tensor(torch.finfo(attn_weights.dtype).min))
-
-    # upcast attention to fp32
-    attn_weights = \
-        torch.nn.functional.softmax(attn_weights,
-                                    dim=-1,
-                                    dtype=torch.float32).to(query_states.dtype)
-    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)
-    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