LLM: add kv_cache optimization for chatglm2-6b-32k (#9165)

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

@@ -182,7 +182,15 @@ def optimize(model):
         pass
 
     if model.config.architectures[0] == "ChatGLMModel":
-        if hasattr(model.config, "padded_vocab_size") and model.config.padded_vocab_size == 65024:
+        if model.config.num_layers == 28 and hasattr(model.config, 'rope_ratio'):
+            # chatglm2-6b-32k
+            modeling_module_name = model.__class__.__module__
+            module = importlib.import_module(modeling_module_name)
+            from bigdl.llm.transformers.models.chatglm2_32k import chatglm2_32k_attention_forward
+            convert_forward(model,
+                            module.SelfAttention,
+                            chatglm2_32k_attention_forward)
+        elif model.config.padded_vocab_size == 65024:
             # chatglm2-6b
             modeling_module_name = model.__class__.__module__
             module = importlib.import_module(modeling_module_name)

python/llm/src/bigdl/llm/transformers/models/chatglm2_32k.py

@@ -0,0 +1,204 @@
+#
+# 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.
+#
+# This file is adapted from
+# https://huggingface.co/THUDM/chatglm2-6b-32k/blob/main/configuration_chatglm.py
+#
+
+import torch
+from typing import Optional, Tuple, Union, List, Callable, Dict, Any
+import torch.nn.functional as F
+from bigdl.llm.transformers.models.utils import init_kv_cache, extend_kv_cache, append_kv_cache
+
+
+KV_CACHE_ALLOC_BLOCK_LENGTH = 256
+KV_CACHE_ALLOC_MIN_LENGTH = 512
+
+
+def split_tensor_along_last_dim(
+        tensor: torch.Tensor,
+        num_partitions: int,
+        contiguous_split_chunks: bool = False,
+) -> List[torch.Tensor]:
+    """Split a tensor along its last dimension.
+    Arguments:
+        tensor: input tensor.
+        num_partitions: number of partitions to split the tensor
+        contiguous_split_chunks: If True, make each chunk contiguous
+                                 in memory.
+    Returns:
+        A list of Tensors
+    """
+    # Get the size and dimension.
+    last_dim = tensor.dim() - 1
+    last_dim_size = tensor.size()[last_dim] // num_partitions
+    # Split.
+    tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
+    # Note: torch.split does not create contiguous tensors by default.
+    if contiguous_split_chunks:
+        return tuple(chunk.contiguous() for chunk in tensor_list)
+
+    return tensor_list
+
+
+@torch.jit.script
+def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
+    # x: [sq, b, np, hn]
+    sq, b, np, hn = x.size(0), x.size(1), x.size(2), x.size(3)
+    rot_dim = rope_cache.shape[-2] * 2
+    x, x_pass = x[..., :rot_dim], x[..., rot_dim:]
+    # truncate to support variable sizes
+    rope_cache = rope_cache[:sq]
+    xshaped = x.reshape(sq, -1, np, rot_dim // 2, 2)
+    rope_cache = rope_cache.view(sq, -1, 1, xshaped.size(3), 2)
+    x_out2 = torch.stack(
+        [
+            xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1],
+            xshaped[..., 1] * rope_cache[..., 0] + xshaped[..., 0] * rope_cache[..., 1],
+        ],
+        -1,
+    )
+    x_out2 = x_out2.flatten(3)
+    return torch.cat((x_out2, x_pass), dim=-1)
+
+
+def chatglm2_32k_attention_forward(
+    self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True
+):
+    # hidden_states: [sq, b, h]
+
+    # =================================================
+    # Pre-allocate memory for key-values for inference.
+    # =================================================
+    # =====================
+    # Query, Key, and Value
+    # =====================
+
+    # Attention heads [sq, b, h] --> [sq, b, (np * 3 * hn)]
+    device = hidden_states.device
+    mixed_x_layer = self.query_key_value(hidden_states)
+
+    if self.multi_query_attention:
+        (query_layer, key_layer, value_layer) = mixed_x_layer.split(
+            [
+                self.num_attention_heads_per_partition * self.hidden_size_per_attention_head,
+                self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
+                self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head,
+            ],
+            dim=-1,
+        )
+        query_layer = query_layer.view(
+            query_layer.size()[:-1] + (self.num_attention_heads_per_partition,
+                                       self.hidden_size_per_attention_head)
+        )
+        key_layer = key_layer.view(
+            key_layer.size()[:-1] + (self.num_multi_query_groups_per_partition,
+                                     self.hidden_size_per_attention_head)
+        )
+        value_layer = value_layer.view(
+            value_layer.size()[:-1]
+            + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
+        )
+    else:
+        new_tensor_shape = mixed_x_layer.size()[:-1] + (self.num_attention_heads_per_partition,
+                                                        3 * self.hidden_size_per_attention_head)
+        mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
+
+        # [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
+        (query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3)
+
+    # apply relative positional encoding (rotary embedding)
+    if rotary_pos_emb is not None:
+        query_layer = apply_rotary_pos_emb(query_layer, rotary_pos_emb)
+        key_layer = apply_rotary_pos_emb(key_layer, rotary_pos_emb)
+
+    cur_length, batch_size = query_layer.shape[0], query_layer.shape[1]
+
+    if self.multi_query_attention:
+        key_length = key_layer.size(0)
+        query_group_size = self.num_attention_heads_per_partition // \
+            self.num_multi_query_groups_per_partition
+        key_layer = key_layer.permute(1, 2, 0, 3).unsqueeze(-3)  # [bs, nh/k, sl, hn]
+        key_layer = key_layer.expand(-1, -1, query_group_size, -1, -1)
+        key_layer = key_layer.contiguous().view((batch_size,
+                                                 self.num_attention_heads_per_partition,
+                                                 key_length,
+                                                 self.hidden_size_per_attention_head))
+        value_layer = value_layer.permute(1, 2, 0, 3).unsqueeze(-3)
+        value_layer = value_layer.expand(-1, -1, query_group_size, -1, -1)
+        value_layer = value_layer.contiguous().view((batch_size,
+                                                     self.num_attention_heads_per_partition,
+                                                     key_length,
+                                                     self.hidden_size_per_attention_head))
+
+    # adjust key and value for inference
+    if kv_cache is not None:
+        cache_k, cache_v = kv_cache
+        cache_k = cache_k.permute(1, 2, 0, 3)
+        cache_v = cache_v.permute(1, 2, 0, 3)
+        past_length = cache_k.size(2)
+
+        if cache_k.stride()[1] <= cache_k.size(2) * cache_k.size(3):
+            max_cache_length = past_length + cur_length + KV_CACHE_ALLOC_BLOCK_LENGTH
+            new_cache_k, new_cache_v = extend_kv_cache(batch_size,
+                                                       self.num_attention_heads_per_partition,
+                                                       self.hidden_size_per_attention_head,
+                                                       past_length,
+                                                       max_cache_length,
+                                                       dtype=query_layer.dtype,
+                                                       device=device)
+            new_cache_k[:] = cache_k
+            new_cache_v[:] = cache_v
+            cache_k = new_cache_k
+            cache_v = new_cache_v
+
+        key_layer, value_layer = append_kv_cache(cache_k, cache_v, key_layer, value_layer)
+    elif use_cache:
+        max_cache_length = max(KV_CACHE_ALLOC_MIN_LENGTH, cur_length) \
+            + KV_CACHE_ALLOC_BLOCK_LENGTH
+        key_cache, value_cache = init_kv_cache(batch_size, self.num_attention_heads_per_partition,
+                                               self.hidden_size_per_attention_head, cur_length,
+                                               max_cache_length,
+                                               dtype=query_layer.dtype, device=device)
+        key_cache[:] = key_layer
+        value_cache[:] = value_layer
+        key_layer = key_cache
+        value_layer = value_cache
+
+    key_layer = key_layer.permute(2, 0, 1, 3)
+    value_layer = value_layer.permute(2, 0, 1, 3)
+
+    if use_cache:
+        if kv_cache is None:
+            kv_cache = torch.cat((key_layer.unsqueeze(0).unsqueeze(0),
+                                  value_layer.unsqueeze(0).unsqueeze(0)), dim=1)
+        else:
+            kv_cache = (key_layer, value_layer)
+    else:
+        kv_cache = None
+
+    # ==================================
+    # core attention computation
+    # ==================================
+
+    context_layer = self.core_attention(query_layer, key_layer, value_layer, attention_mask)
+
+    # =================
+    # Output. [sq, b, h]
+    # =================
+
+    output = self.dense(context_layer)
+
+    return output, kv_cache