groupuser/phi-1_5
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Upload modeling_phi.py

Browse Source
This commit is contained in:
Gustavo de Rosa 2024-01-10 13:54:40 +00:00 committed by huggingface-web
parent 3a705a2d6b
commit 914c8fb3c6
1 changed files with 12 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
modeling_phi.py
View File

@ -358,7 +358,9 @@ class PhiAttention(nn.Module):
key_states = repeat_kv(key_states, self.num_key_value_groups) key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups)
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim) attn_weights = torch.matmul(
query_states.to(torch.float32), key_states.to(torch.float32).transpose(2, 3)
) / math.sqrt(self.head_dim)
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len): if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
raise ValueError( raise ValueError(
@ -374,7 +376,7 @@ class PhiAttention(nn.Module):
attn_weights = attn_weights + attention_mask attn_weights = attn_weights + attention_mask
# upcast attention to fp32 # upcast attention to fp32
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training) attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
attn_output = torch.matmul(attn_weights, value_states) attn_output = torch.matmul(attn_weights, value_states)
@ -483,8 +485,10 @@ class PhiFlashAttention2(PhiAttention):
# in fp32. # in fp32.
if query_states.dtype == torch.float32: if query_states.dtype == torch.float32:
if torch.is_autocast_enabled():
target_dtype = torch.get_autocast_gpu_dtype()
# Handle the case where the model is quantized # Handle the case where the model is quantized
if hasattr(self.config, "_pre_quantization_dtype"): elif hasattr(self.config, "_pre_quantization_dtype"):
target_dtype = self.config._pre_quantization_dtype target_dtype = self.config._pre_quantization_dtype
else: else:
target_dtype = self.q_proj.weight.dtype target_dtype = self.q_proj.weight.dtype
@ -1093,7 +1097,7 @@ class PhiForCausalLM(PhiPreTrainedModel):
# Keep only the unprocessed tokens: # Keep only the unprocessed tokens:
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
# some of the inputs are exclusivelly passed as part of the cache (e.g. when passing input_embeds as # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
# input) # input)
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]: if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :] input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
@ -1225,9 +1229,10 @@ class PhiForSequenceClassification(PhiPreTrainedModel):
sequence_lengths = -1 sequence_lengths = -1
else: else:
if input_ids is not None: if input_ids is not None:
sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to( # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
logits.device sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
) sequence_lengths = sequence_lengths % input_ids.shape[-1]
sequence_lengths = sequence_lengths.to(logits.device)
else: else:
sequence_lengths = -1 sequence_lengths = -1