LLama optimize_model to support transformers 4.36 (#9818)

* supoort 4.36

* style

* update

* update

* update

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

@@ -432,10 +432,6 @@ def _optimize_post(model, lightweight_bmm=False):
 
     trans_version = transformers.__version__
     if version.parse(trans_version) >= version.parse("4.31.0"):
-        convert_forward(
-            model,
-            transformers.models.llama.modeling_llama.LlamaAttention,
-            llama_attention_forward_4_31,)
         convert_forward(
             model,
             transformers.models.llama.modeling_llama.LlamaRMSNorm,
@@ -443,6 +439,19 @@ def _optimize_post(model, lightweight_bmm=False):
         convert_forward(model,
                         transformers.models.llama.modeling_llama.LlamaMLP,
                         llama_mlp_forward)
+        if version.parse(trans_version) >= version.parse("4.36.0"):
+            # transformers version >= 4.36.0
+            from bigdl.llm.transformers.models.llama import llama_attention_forward_4_36
+            convert_forward(
+                model,
+                transformers.models.llama.modeling_llama.LlamaAttention,
+                llama_attention_forward_4_36, )
+        else:
+            # transformers version between 4.31.0 - 4.35.2
+            convert_forward(
+                model,
+                transformers.models.llama.modeling_llama.LlamaAttention,
+                llama_attention_forward_4_31, )
             if enable_vllm_se_batching:
                 convert_forward(
                     model,

python/llm/src/bigdl/llm/transformers/models/llama.py

@@ -32,15 +32,16 @@
 # limitations under the License.
 
 import torch
+import warnings
 import importlib
 import torch.nn as nn
 from typing import Optional, Tuple, Union, List
 import math
 import os
 import torch.nn.functional as F
-from bigdl.llm.utils.common import invalidInputError
 from bigdl.llm.transformers.models.utils import init_kv_cache, extend_kv_cache, append_kv_cache
-from bigdl.llm.transformers.models.utils import is_enough_kv_cache_room_4_31, apply_rotary_pos_emb
+from bigdl.llm.transformers.models.utils import is_enough_kv_cache_room_4_31, \
+    apply_rotary_pos_emb, is_enough_kv_cache_room_4_36
 from bigdl.llm.transformers.models.utils import apply_rotary_pos_emb_no_cache_xpu
 from bigdl.llm.transformers.models.utils import use_flash_attention, use_esimd_sdp
 from transformers.modeling_outputs import BaseModelOutputWithPast
@@ -510,6 +511,210 @@ def llama_attention_selective_batching_forward_4_31(
     return attn_output.to(original_dtype), attn_weights, updated_past_key_values
 
 
+def llama_attention_forward_4_36(
+    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.`"
+        )
+
+    bsz, q_len, _ = hidden_states.size()
+    device = hidden_states.device
+    # for flash attention
+    original_dtype = hidden_states.dtype
+    if not self.training and not hidden_states.requires_grad:
+        fsdp_flag = use_flash_attention(hidden_states)
+    else:
+        fsdp_flag = False
+    if fsdp_flag:
+        attention_dtype = torch.float16  # use fp16 for flash attention
+    else:
+        attention_dtype = original_dtype
+
+    use_fuse_rope = should_use_fuse_rope(self, hidden_states, position_ids)
+    enough_kv_room = is_enough_kv_cache_room_4_36(past_key_value, self.layer_idx)
+    qtype = getattr(self.q_proj, "qtype", None)
+    is_q4_0 = qtype == SYM_INT4
+    no_tp = not self.config.pretraining_tp > 1
+    decoding_fast_path = (no_tp and is_q4_0 and use_fuse_rope and
+                          enough_kv_room and bsz * q_len == 1)
+
+    # single batch decoding fast path
+    # forward_qkv takes will perform QKV projection, rotary position embedding
+    # and save the key/value states to cache, then return query states and the
+    # extended key/value cache
+    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
+        query_states, key_states, value_states = linear_q4_0.forward_qkv(hidden_states,
+                                                                         self.q_proj.weight,
+                                                                         self.k_proj.weight,
+                                                                         self.v_proj.weight,
+                                                                         position_ids,
+                                                                         cache_k, cache_v,
+                                                                         self.q_proj.weight.qtype,
+                                                                         kv_seq_len,
+                                                                         self.head_dim)
+        kv_seq_len += 1
+        # update past_key_value's seem_tokens and kv caches.
+        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:
+        if self.config.pretraining_tp > 1:
+            key_value_slicing = ((self.num_key_value_heads * self.head_dim) //
+                                 self.config.pretraining_tp)
+            query_slices = self.q_proj.weight.split((self.num_heads * self.head_dim)
+                                                    // self.config.pretraining_tp, dim=0)
+            key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
+            value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
+
+            query_states = [F.linear(hidden_states, query_slices[i])
+                            for i in range(self.config.pretraining_tp)]
+            query_states = torch.cat(query_states, dim=-1)
+
+            key_states = [F.linear(hidden_states, key_slices[i])
+                          for i in range(self.config.pretraining_tp)]
+            key_states = torch.cat(key_states, dim=-1)
+
+            value_states = [F.linear(hidden_states, value_slices[i])
+                            for i in range(self.config.pretraining_tp)]
+            value_states = torch.cat(value_states, dim=-1)
+        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 decodingwith 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)
+
+        if use_fuse_rope:
+            query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states,
+                                                                         key_states,
+                                                                         position_ids,
+                                                                         "llama")
+        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, "llama")
+
+        if past_key_value is not None:
+            # update the number of seen tokens
+            if self.layer_idx == 0:
+                past_key_value.seen_tokens += key_states.shape[-2]
+
+            # reuse k, v, self_attention
+            # update `past_key_value` with `key_states` and `value_states` for layer `layer_idx`
+            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).to(device,
+                                                                     dtype=attention_dtype)
+    value_states = repeat_kv(value_states, self.num_key_value_groups).to(device,
+                                                                         dtype=attention_dtype)
+
+    if fsdp_flag:
+        # now only use flash attention for first token
+        attn_output = F.scaled_dot_product_attention(query_states.to(dtype=attention_dtype),
+                                                     key_states,
+                                                     value_states,
+                                                     is_causal=True)
+        attn_weights = None
+    elif use_esimd_sdp(q_len, self.head_dim, query_states):
+        import linear_fp16_esimd
+        attn_output = linear_fp16_esimd.sdp_forward(query_states,
+                                                    key_states.contiguous(),
+                                                    value_states.contiguous())
+        attn_output = attn_output.view(query_states.shape)
+        attn_weights = None
+    else:
+        # otherwise, use native attention
+        attn_output, attn_weights = native_sdp(query_states, key_states, value_states,
+                                               attention_mask,
+                                               bsz, q_len, kv_seq_len,
+                                               self.head_dim, self.num_heads)
+
+    attn_output_size = (bsz, self.num_heads, q_len, self.head_dim)
+    if attn_output.size() != attn_output_size:
+        invalidInputError(False,
+                          f"`attn_output` should be of size {attn_output_size},"
+                          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)
+
+    if self.config.pretraining_tp > 1:
+        attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
+        o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp,
+                                                 dim=1)
+        attn_output = sum([F.linear(attn_output[i], o_proj_slices[i])
+                           for i in range(self.config.pretraining_tp)])
+    else:
+        attn_output = self.o_proj(attn_output)
+
+    if not output_attentions:
+        attn_weights = None
+
+    return attn_output.to(original_dtype), attn_weights, past_key_value
+
+
 def native_sdp(query, key, value, attention_mask,
                bsz, q_len, kv_seq_len, head_dim, num_heads):
     attn_weights = torch.matmul(query,