ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

LLM: update qwen attention forward. (#9695)

Browse source 
* feat: update qwen attention forward.

* fix: style.
This commit is contained in:
Cengguang Zhang 2023-12-15 14:06:15 +08:00 committed by GitHub
parent b8437a1c1e
commit adbef56001
1 changed files with 36 additions and 14 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

50
python/llm/src/bigdl/llm/transformers/models/qwen.py
View file

@ -14,7 +14,7 @@
# limitations under the License. # limitations under the License.
# #
# Some parts of this file is adapted from # Some parts of this file is adapted from
# https://huggingface.co/Qwen/Qwen-7B-Chat/blob/faf3ff60438d724a7eb78ebed7e2f7c7330c6bd8/modeling_qwen.py # https://huggingface.co/Qwen/Qwen-7B-Chat/blob/be72f02dd47087f9035ee9bb5dea571b84785d27/modeling_qwen.py
# #
# Copyright (c) Alibaba Cloud. # Copyright (c) Alibaba Cloud.
# #
@ -38,7 +38,7 @@ except ImportError:
from bigdl.llm.transformers.models.utils import extend_kv_cache, init_kv_cache, append_kv_cache from bigdl.llm.transformers.models.utils import extend_kv_cache, init_kv_cache, append_kv_cache
from bigdl.llm.transformers.models.utils import rotate_half from bigdl.llm.transformers.models.utils import rotate_half
from bigdl.llm.utils.common import invalidInputError from bigdl.llm.utils.common import invalidInputError, invalidOperationError
from bigdl.llm.ggml.quantize import ggml_tensor_qtype from bigdl.llm.ggml.quantize import ggml_tensor_qtype
apply_rotary_emb_func = None apply_rotary_emb_func = None
@ -48,6 +48,7 @@ flash_attn_unpadded_func = None
logger = logging.get_logger(__name__) logger = logging.get_logger(__name__)
KV_CACHE_ALLOC_BLOCK_LENGTH = 256 KV_CACHE_ALLOC_BLOCK_LENGTH = 256
SUPPORT_TORCH2 = hasattr(torch, '__version__') and int(torch.__version__.split(".")[0]) >= 2
def apply_rotary_pos_emb(t, freqs): def apply_rotary_pos_emb(t, freqs):
@ -159,7 +160,7 @@ def qwen_attention_forward(
else: else:
seq_start = key.size(1) - query.size(1) seq_start = key.size(1) - query.size(1)
seq_end = key.size(1) seq_end = key.size(1)
logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :] logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :].type_as(query)
query = query * logn_tensor.expand_as(query) query = query * logn_tensor.expand_as(query)
if ( if (
@ -169,12 +170,12 @@ def qwen_attention_forward(
and query.is_cuda and query.is_cuda
): ):
q, k, v = query, key, value q, k, v = query, key, value
context_layer = self.core_attention_flash(q, k, v, attention_mask=attention_mask) attn_output = self.core_attention_flash(q, k, v, attention_mask=attention_mask)
else: else:
key_size = key[0].size(2) if self.use_cache_quantization else key.size(1)
if query.size(1) == key_size: if query.size(1) == key_size:
causal_mask = torch.tril( causal_mask = torch.tril(
torch.ones((key_size, key_size), dtype=torch.bool, device=key.device) torch.ones((key_size, key_size), dtype=torch.bool, device=query.device)
).view(1, 1, key_size, key_size) ).view(1, 1, key_size, key_size)
else: else:
causal_mask = None causal_mask = None
@ -189,13 +190,30 @@ def qwen_attention_forward(
and not self.is_fp32 and not self.is_fp32
and not query.is_cuda and not query.is_cuda
): ):
invalidInputError(False, _ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED) invalidOperationError(False,
attn_output, attn_weight = self._attn( None,
query, key, value, causal_mask, attention_mask, head_mask None,
) Exception(_ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED))
context_layer = self._merge_heads(
attn_output, self.num_heads, self.head_dim if not self.use_cache_quantization and SUPPORT_TORCH2:
) if attention_mask is not None:
attention_mask = attention_mask.expand(-1, -1, query.size(2), -1)
if causal_mask is not None:
attention_mask = attention_mask.masked_fill(~causal_mask,
torch.finfo(query.dtype).min)
else:
attention_mask = causal_mask
attn_output = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask
).transpose(1, 2)
attn_weight = None
else:
attn_output, attn_weight = self._attn(
query, key, value, causal_mask, attention_mask, head_mask
)
context_layer = self._merge_heads(
attn_output, self.num_heads, self.head_dim
)
attn_output = self.c_proj(context_layer) attn_output = self.c_proj(context_layer)
@ -206,7 +224,11 @@ def qwen_attention_forward(
and flash_attn_unpadded_func is not None and flash_attn_unpadded_func is not None
and not self.is_fp32 and not self.is_fp32
): ):
invalidInputError(False, "Cannot output attentions while using flash-attn") invalidInputError(False,
f"Cannot output attentions while using flash-attn")
elif not self.use_cache_quantization and SUPPORT_TORCH2:
invalidInputError(False,
f"Cannot output attentions while using scaled_dot_product_attention")
else: else:
outputs += (attn_weight,) outputs += (attn_weight,)