refactor baichuan2-7b (#11062)

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

@@ -710,6 +710,12 @@ def _optimize_pre(model):
         model.apply(pre_compute_inv_freq)
         from ipex_llm.transformers.models.phi3 import split_mlp
         model.apply(split_mlp)
+    # for baichuan2
+    if model.config.model_type == "baichuan" and model.config.vocab_size == 125696:
+        if model.config.hidden_size in [4096, 2048]:
+            # baichuan2-7B
+            from ipex_llm.transformers.models.baichuan2 import pre_compute_inv_freq
+            model.apply(pre_compute_inv_freq)
     if model.config.model_type == "qwen":
         rope_base = model.config.rotary_emb_base
         from accelerate.big_modeling import init_empty_weights

python/llm/src/ipex_llm/transformers/models/baichuan2.py

@@ -23,50 +23,30 @@ from typing import Optional, Tuple
 import torch
 import torch.utils.checkpoint
 from torch.nn import functional as F
-from ipex_llm.ggml.quantize import ggml_tensor_qtype
 from ipex_llm.transformers.models.utils import init_fp8_kv_cache, append_fp8_kv_cache, \
     restore_fp8_kv_cache, use_quantize_kv_cache
 from ipex_llm.transformers.models.utils import init_kv_cache, extend_kv_cache, \
     append_kv_cache, is_enough_kv_cache_room_4_31
-from ipex_llm.transformers.models.utils import use_flash_attention, use_sdp
+from ipex_llm.transformers.models.utils import should_use_fuse_rope
+from ipex_llm.transformers.models.utils import use_flash_attention, use_sdp, use_sdp_causal
 from ipex_llm.transformers.models.utils import apply_rotary_pos_emb, SILU
-from ipex_llm.transformers.models.utils import apply_rotary_pos_emb_no_cache_xpu
 from ipex_llm.transformers.models.utils import mlp_fusion_check
-from ipex_llm.utils.common.log4Error import invalidInputError
+import warnings
-from transformers.utils import logging
-logger = logging.get_logger(__name__)
-
-try:
-    from xformers import ops as xops
-except ImportError:
-    xops = None
-    logger.warning(
-        "Xformers is not installed correctly. If you want to use memory_efficient_attention to "
-        "accelerate training use the following command to install Xformers\npip install xformers."
-    )
-
 import os
 
+
 KV_CACHE_ALLOC_BLOCK_LENGTH = int(os.environ.get("KV_CACHE_ALLOC_BLOCK_LENGTH", 256))
 
 
-def should_split_qkv_tensor(query_states, bsz, num_heads, q_len, kv_seq_len, output_attentions):
+def pre_compute_inv_freq(module: torch.nn.Module):
-    if not output_attentions:
+    if module.__class__.__name__ == "RotaryEmbedding":
-        if os.environ.get("IPEX_LLM_SPLIT_QKV", None) is not None:
+        inv_freq = module.inv_freq
-            return os.environ.get("IPEX_LLM_SPLIT_QKV", None) == "1"
+        del module.inv_freq
-        elif query_states.dtype == torch.float16 and \
+        module.register_buffer("inv_freq", inv_freq, persistent=False)
-                query_states.shape[2] >= 5400:
-            # split tensor for memory block limitation
-            # support fp16 and set input length threshold at 5400 for now
-            return True
-        elif query_states.element_size()*bsz*num_heads*q_len*kv_seq_len >= 4*1024**3:
-            # attn_weight size larger than memory block limitation 4GB
-            return True
-    return False
 
 
 def baichuan_13b_rms_norm_forward(self, hidden_states):
-    if hidden_states.device.type == "xpu" and not (self.training and hidden_states.requires_grad):
+    if hidden_states.device.type == "xpu" and not (self.training or hidden_states.requires_grad):
         import linear_q4_0
         x_2d = hidden_states.reshape(-1, hidden_states.size(-1)).contiguous()
         output = linear_q4_0.rms_norm(self.weight, x_2d, self.epsilon)
@@ -105,57 +85,37 @@ def baichuan_attention_forward_7b(
     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.W_pack, hidden_states):
-        forward_function = baichuan_attention_forward_7b_quantized
-    else:
-        forward_function = baichuan_attention_forward_7b_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 baichuan_attention_forward_7b_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
 
-    proj = self.W_pack(hidden_states)
+    qkv = self.W_pack(hidden_states)
-    proj = torch.chunk(proj, 3, -1)
+    qkv = qkv.view(bsz, q_len, self.num_heads * 3, self.head_dim)
-    # batch_size x source_len x hidden_size
+    qkv = qkv.transpose(1, 2)
-    query_states = proj[0].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+    query_states, key_states, value_states = qkv.split([self.num_heads,
-    # batch_size x target_len x head_size
+                                                        self.num_heads,
-    key_states = proj[1].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+                                                        self.num_heads], dim=1)
-    # batch_size x source_len x hidden_size
-    value_states = proj[2].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
 
-    kv_seq_len = key_states.shape[-2]
+    kv_seq_len = key_states.shape[2]
     if past_key_value is not None:
-        kv_seq_len += past_key_value[0].shape[-2]
+        kv_seq_len += past_key_value[0].shape[2]
-    if query_states.device.type == "xpu" and not (self.training and query_states.requires_grad):
+
-        query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states,
+    # IPEX-LLM OPT: fuse rope
-                                                                     key_states,
+    if should_use_fuse_rope(hidden_states, position_ids, self.training):
-                                                                     position_ids,
+        import linear_q4_0
-                                                                     "baichuan")
+        linear_q4_0.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids,
+                                         query_states, key_states)
     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, "baichuan")
+        query_states = query_states.to(hidden_states.dtype)
+        key_states = key_states.to(hidden_states.dtype)
+
+    # IPEX-LLM OPT: kv cache and quantize kv
+    use_quantize_kv = use_quantize_kv_cache(self.W_pack, hidden_states)
+    if use_quantize_kv:
         if past_key_value is None:
-        kv_seq_len = key_states.shape[-2]
             k_cache, v_cache = init_fp8_kv_cache(
                 bsz, self.num_heads, kv_seq_len, self.head_dim,
                 device=device
@@ -164,170 +124,84 @@ def baichuan_attention_forward_7b_quantized(
             k_cache, v_cache = past_key_value
         key_states, value_states = append_fp8_kv_cache(k_cache, v_cache,
                                                        key_states, value_states)
-
-    past_key_value = (key_states, value_states) if use_cache else None
-
-    invalidInputError(attention_mask is None or attention_mask.dtype != torch.bool,
-                      "attention_mask's dtype cannot be bool")
-
-    scaling_factor = 1 / math.sqrt(query_states.size(-1))
-    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)
-        if should_split_qkv_tensor(query_states, bsz, self.num_heads,
-                                   q_len, kv_seq_len, output_attentions):
-            attn_output, attn_weights = native_sdp_split_qkv_tensor(query_states, key_states,
-                                                                    value_states, attention_mask)
     else:
-            attn_output = torch.matmul(query_states * scaling_factor, key_states.transpose(-2, -1))
+        if past_key_value is None:
-
-            if attention_mask is not None:
-                attn_output += attention_mask
-            attn_output = torch.softmax(attn_output, -1)
-            attn_output = attn_output.to(hidden_states.dtype)
-            attn_output = torch.matmul(attn_output, value_states)
-    else:
-        import linear_q4_0
-        attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states,
-                                          attention_mask)
-        attn_weights = None
-    attn_output = attn_output.transpose(1, 2)
-
-    attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim)
-    attn_output = self.o_proj(attn_output)
-
-    if not output_attentions:
-        attn_weights = None
-
-    return attn_output, attn_weights, past_key_value
-
-
-def baichuan_attention_forward_7b_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]]]:
-    bsz, q_len, _ = hidden_states.size()
-    device = hidden_states.device
-
-    proj = self.W_pack(hidden_states)
-    proj = torch.chunk(proj, 3, -1)
-    # batch_size x source_len x hidden_size
-    query_states = proj[0].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-    # batch_size x target_len x head_size
-    key_states = proj[1].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-    # batch_size x source_len x hidden_size
-    value_states = proj[2].view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-
-    kv_seq_len = key_states.shape[-2]
-    enough_kv_room = True
-    if past_key_value is not None:
-        enough_kv_room = is_enough_kv_cache_room_4_31(past_key_value, seq_len=kv_seq_len)
-        kv_seq_len += past_key_value[0].shape[-2]
-    if query_states.device.type == "xpu" and not (self.training and query_states.requires_grad):
-        query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states,
-                                                                     key_states,
-                                                                     position_ids,
-                                                                     "baichuan")
-    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, "baichuan")
-    # [bsz, nh, t, hd]
-
-    # if past_key_value is not None:
-    #     # reuse k, v, self_attention
-    #     key_states = torch.cat([past_key_value[0], key_states], dim=2)
-    #     value_states = torch.cat([past_key_value[1], value_states], dim=2)
-    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,
+            k_cache, v_cache = 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
+            k_cache[...] = key_states
-        new_value_states[:] = value_states
+            v_cache[...] = value_states
-        key_states = new_key_states
+            key_states = k_cache
-        value_states = new_value_states
+            value_states = v_cache
+        else:
+            k_cache, v_cache = past_key_value
+            if k_cache.stride(1) < kv_seq_len * k_cache.size(3):
+                max_cache_length = kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH
+                new_k_cache, new_v_cache = extend_kv_cache(bsz,
+                                                           self.num_heads,
+                                                           self.head_dim,
+                                                           k_cache.size(2),
+                                                           max_cache_length,
+                                                           dtype=k_cache.dtype,
+                                                           device=device)
+                new_k_cache[...] = k_cache
+                new_v_cache[...] = v_cache
+                k_cache = new_k_cache
+                v_cache = new_v_cache
+            key_states, value_states = append_kv_cache(k_cache, v_cache, key_states, value_states)
 
     past_key_value = (key_states, value_states) if use_cache else None
 
-    invalidInputError(attention_mask is None or attention_mask.dtype != torch.bool,
+    if self.training:
-                      "attention_mask's dtype cannot be bool")
+        warnings.warn("xops is not supported on Intel GPU, so just use normal implementation")
 
-    if xops is not None and self.training:
     attn_weights = None
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.transpose(1, 2)
-        attn_output = xops.memory_efficient_attention(
-            query_states, key_states, value_states, attn_bias=xops.LowerTriangularMask()
-        )
-    else:
     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(dtype=torch.float16),
                                                      key_states.to(dtype=torch.float16),
                                                      value_states.to(dtype=torch.float16),
-                                                         is_causal=True)
+                                                     is_causal=True).to(hidden_states.dtype)
-            attn_weights = None
+    elif use_sdp(q_len, kv_seq_len, self.head_dim, query_states):
-        elif not self.training and not hidden_states.requires_grad and \
-                use_sdp(q_len, key_states.shape[2], self.head_dim, query_states):
         import linear_q4_0
-            attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask)
+        if use_quantize_kv:
-            attn_output = attn_output.view(query_states.shape)
+            attn_output = linear_q4_0.sdp_fp8(query_states, key_states, value_states,
-            attn_weights = None
-        else:
-            if should_split_qkv_tensor(query_states, bsz, self.num_heads,
-                                       q_len, kv_seq_len, output_attentions):
-                attn_output, attn_weights = native_sdp_split_qkv_tensor(query_states,
-                                                                        key_states,
-                                                                        value_states,
                                               attention_mask)
         else:
-                scaling_factor = 1 / math.sqrt(query_states.size(-1))
+            attn_output = linear_q4_0.sdp(query_states, key_states, value_states, attention_mask)
-                attn_output = torch.matmul(query_states * scaling_factor,
+    elif use_sdp_causal(q_len, kv_seq_len, self.head_dim, query_states, self.training):
-                                           key_states.transpose(-2, -1))
+        import linear_q4_0
+        if use_quantize_kv:
+            attn_output = linear_q4_0.sdp_fp8_causal(query_states, key_states, value_states)
+        else:
+            attn_output = linear_q4_0.sdp_causal(query_states, key_states, value_states)
+    else:
+        if use_quantize_kv:
+            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)) / math.sqrt(self.head_dim)
         if attention_mask is not None:
-                    attn_output += attention_mask
+            attn_weights = attn_weights + attention_mask
-                attn_output = torch.softmax(attn_output, -1)
+        # upcast attention to fp32
-                attn_output = torch.matmul(attn_output, value_states)
+        attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1,
+                                                   dtype=torch.float32).to(value_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
 
-        attn_output = attn_output.transpose(1, 2)
-    attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim)
     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
+    return attn_output, attn_weights, past_key_value
 
 
 def baichuan_attention_forward_13b(
@@ -507,20 +381,10 @@ def baichuan_attention_forward_13b_origin(
         value_states = new_value_states
 
     past_key_value = (key_states, value_states) if use_cache else None
-    if xops is not None and self.training:
+
-        attn_weights = None
+    if self.training:
-        # query_states = query_states.transpose(1, 2)
+        warnings.warn("xops is not supported on Intel GPU, so just use normal implementation")
-        # key_states = key_states.transpose(1, 2)
+
-        # value_states = value_states.transpose(1, 2)
-        # attn_output = xops.memory_efficient_attention(
-        #     query_states, key_states, value_states, attn_bias=attention_mask
-        # )
-        with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True,
-                                            enable_mem_efficient=True):
-            attn_output = F.scaled_dot_product_attention(query_states, key_states, value_states,
-                                                         attn_mask=attention_mask)
-        attn_output = attn_output.transpose(1, 2)
-    else:
     attn_weights = torch.matmul(
         query_states.to(dtype=key_states.dtype), key_states.transpose(2, 3)
     ) / math.sqrt(self.head_dim)
@@ -647,21 +511,3 @@ def baichuan_13b_get_alibi_mask(self, tensor, seq_length_with_past):
             : self.n_head, :seq_length_with_past, :seq_length_with_past
         ]
     return mask
-
-
-def native_sdp_split_qkv_tensor(query, key, value, attention_mask):
-    block_size = 8
-    query_split = torch.split(query, block_size, dim=1)
-    key_split = torch.split(key.transpose(-2, -1), block_size, dim=1)
-    value_split = torch.split(value, block_size, dim=1)
-    attn_outputs = []
-    scaling_factor = 1 / math.sqrt(query.size(-1))
-    for q, k, v in zip(query_split, key_split, value_split):
-        attn_output_split = torch.matmul(q * scaling_factor, k)
-        if attention_mask is not None:
-            attn_output_split += attention_mask
-        attn_output_split = torch.softmax(attn_output_split, -1)
-        attn_output_split = torch.matmul(attn_output_split, v)
-        attn_outputs.append(attn_output_split)
-    attn_output = torch.cat(attn_outputs, dim=1)
-    return attn_output, None