fix(server): fix escape characters in stop sequence (#155)

This commit is contained in:
OlivierDehaene 2023-04-05 19:37:41 +02:00 committed by GitHub
parent 9122e7bd9c
commit 3f2542bb6a
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 90 additions and 69 deletions

View File

@ -14,6 +14,15 @@ def test_stop_sequence_criteria():
assert not criteria("/test; ") assert not criteria("/test; ")
def test_stop_sequence_criteria_escape():
criteria = StopSequenceCriteria("<|stop|>")
assert not criteria("<")
assert not criteria("<|stop")
assert criteria("<|stop|>")
assert not criteria("<|stop|> ")
def test_stopping_criteria(): def test_stopping_criteria():
criteria = StoppingCriteria(0, [StopSequenceCriteria("/test;")], max_new_tokens=5) criteria = StoppingCriteria(0, [StopSequenceCriteria("/test;")], max_new_tokens=5)
assert criteria(65827, "/test") == (False, None) assert criteria(65827, "/test") == (False, None)

View File

@ -47,12 +47,12 @@ class FastLayerNorm(nn.LayerNorm):
class FastLinear(nn.Linear): class FastLinear(nn.Linear):
def __init__( def __init__(
self, self,
in_features: int, in_features: int,
out_features: int, out_features: int,
bias: bool = True, bias: bool = True,
device=None, device=None,
dtype=None, dtype=None,
) -> None: ) -> None:
super(FastLinear, self).__init__(in_features, out_features, bias, device, dtype) super(FastLinear, self).__init__(in_features, out_features, bias, device, dtype)
@ -67,10 +67,10 @@ class FastLinear(nn.Linear):
class FlashMQAttention(torch.nn.Module): class FlashMQAttention(torch.nn.Module):
def __init__( def __init__(
self, self,
num_heads, num_heads,
hidden_size, hidden_size,
process_group=None, process_group=None,
): ):
super().__init__() super().__init__()
self.num_heads = num_heads self.num_heads = num_heads
@ -86,13 +86,13 @@ class FlashMQAttention(torch.nn.Module):
raise NotImplementedError raise NotImplementedError
def forward( def forward(
self, self,
hidden_states, hidden_states,
cu_seqlens, cu_seqlens,
max_s, max_s,
layer_past, layer_past,
layer_past_present_indices, layer_past_present_indices,
cu_seqlens_q, cu_seqlens_q,
): ):
qkv = self.attn(hidden_states) qkv = self.attn(hidden_states)
@ -162,15 +162,17 @@ class FlashMQAttention(torch.nn.Module):
class MLP(nn.Module): class MLP(nn.Module):
def __init__( def __init__(self, act, hidden_size, intermediate_size, process_group=None):
self, act, hidden_size, intermediate_size, process_group=None
):
super().__init__() super().__init__()
self.act = ( self.act = (
ACT2FN[act] ACT2FN[act]
if "gelu" not in act if "gelu" not in act
else lambda x: torch.nn.functional.gelu(x, approximate="tanh" if act in ["gelu_fast", else lambda x: torch.nn.functional.gelu(
"gelu_pytorch_tanh"] else None) x,
approximate="tanh"
if act in ["gelu_fast", "gelu_pytorch_tanh"]
else None,
)
) )
if process_group is None: if process_group is None:
@ -188,13 +190,13 @@ class MLP(nn.Module):
class Block(nn.Module): class Block(nn.Module):
def __init__( def __init__(
self, self,
num_heads, num_heads,
act, act,
hidden_size, hidden_size,
intermediate_size, intermediate_size,
layer_norm_eps, layer_norm_eps,
process_group=None, process_group=None,
): ):
super().__init__() super().__init__()
self.ln_1 = FastLayerNorm(hidden_size, eps=layer_norm_eps) self.ln_1 = FastLayerNorm(hidden_size, eps=layer_norm_eps)
@ -212,14 +214,14 @@ class Block(nn.Module):
) )
def forward( def forward(
self, self,
hidden_states, hidden_states,
residual, residual,
cu_seqlens, cu_seqlens,
max_s, max_s,
layer_past, layer_past,
layer_past_present_indices, layer_past_present_indices,
cu_seqlens_q, cu_seqlens_q,
): ):
hidden_states, residual = self.ln_1(hidden_states, residual) hidden_states, residual = self.ln_1(hidden_states, residual)
@ -232,9 +234,7 @@ class Block(nn.Module):
cu_seqlens_q, cu_seqlens_q,
) )
hidden_states, residual = self.ln_2( hidden_states, residual = self.ln_2(hidden_states, residual)
hidden_states, residual
)
mlp_output = self.mlp(hidden_states) mlp_output = self.mlp(hidden_states)
@ -258,16 +258,16 @@ class FlashSantacoderModel(nn.Module):
config.num_attention_heads, config.num_attention_heads,
config.activation_function, config.activation_function,
config.hidden_size, config.hidden_size,
config.n_inner if config.n_inner is not None else 4 * config.hidden_size, config.n_inner
if config.n_inner is not None
else 4 * config.hidden_size,
config.layer_norm_epsilon, config.layer_norm_epsilon,
process_group, process_group,
) )
for _ in range(config.num_hidden_layers) for _ in range(config.num_hidden_layers)
] ]
) )
self.ln_f = FastLayerNorm( self.ln_f = FastLayerNorm(config.hidden_size, eps=config.layer_norm_epsilon)
config.hidden_size, eps=config.layer_norm_epsilon
)
self.head_size = self.h[0].attn.head_size self.head_size = self.h[0].attn.head_size
self.num_heads = self.h[0].attn.num_heads self.num_heads = self.h[0].attn.num_heads
@ -281,12 +281,12 @@ class FlashSantacoderModel(nn.Module):
layer.mlp.c_proj.transpose_weight() layer.mlp.c_proj.transpose_weight()
def forward( def forward(
self, self,
input_ids, input_ids,
position_ids, position_ids,
cu_seqlens, cu_seqlens,
max_s, max_s,
past_key_values=None, past_key_values=None,
): ):
hidden_states = self.wte(input_ids) + self.wpe(position_ids) hidden_states = self.wte(input_ids) + self.wpe(position_ids)
@ -335,21 +335,19 @@ class FlashSantacoderForCausalLM(nn.Module):
self.transformer = FlashSantacoderModel(config, process_group) self.transformer = FlashSantacoderModel(config, process_group)
self.lm_head = FastLinear( self.lm_head = FastLinear(config.hidden_size, config.vocab_size, bias=False)
config.hidden_size, config.vocab_size, bias=False
)
def post_load_weights(self): def post_load_weights(self):
self.transformer.post_load_weights() self.transformer.post_load_weights()
self.lm_head.transpose_weight() self.lm_head.transpose_weight()
def forward( def forward(
self, self,
input_ids, input_ids,
position_ids, position_ids,
cu_seqlens, cu_seqlens,
max_s, max_s,
past_key_values=None, past_key_values=None,
): ):
hidden_states, present = self.transformer( hidden_states, present = self.transformer(
input_ids, position_ids, cu_seqlens, max_s, past_key_values input_ids, position_ids, cu_seqlens, max_s, past_key_values

View File

@ -9,7 +9,7 @@ from typing import Optional, List
from text_generation_server.models import FlashCausalLM from text_generation_server.models import FlashCausalLM
from text_generation_server.models.custom_modeling.flash_santacoder_modeling import ( from text_generation_server.models.custom_modeling.flash_santacoder_modeling import (
FlashSantacoderForCausalLM FlashSantacoderForCausalLM,
) )
from text_generation_server.utils import ( from text_generation_server.utils import (
weight_files, weight_files,
@ -37,8 +37,9 @@ class FlashSantacoder(FlashCausalLM):
) )
config = AutoConfig.from_pretrained( config = AutoConfig.from_pretrained(
model_id, revision=revision, model_id,
trust_remote_code=True # Needed as the config is not part of Transformers revision=revision,
trust_remote_code=True, # Needed as the config is not part of Transformers
) )
# We do not use from_pretrained as we modified the model internal module layout # We do not use from_pretrained as we modified the model internal module layout
@ -65,8 +66,8 @@ class FlashSantacoder(FlashCausalLM):
@staticmethod @staticmethod
def load_weights( def load_weights(
model: FlashSantacoderForCausalLM, model: FlashSantacoderForCausalLM,
filenames: List[Path], filenames: List[Path],
): ):
for filename in filenames: for filename in filenames:
state_dict = torch.load(filename, map_location="cpu") state_dict = torch.load(filename, map_location="cpu")
@ -91,7 +92,12 @@ class FlashSantacoder(FlashCausalLM):
current_parameter_tensor = None current_parameter_tensor = None
if current_parameter_tensor is not None: if current_parameter_tensor is not None:
if "c_fc.weight" in key or "c_proj.weight" in key or "q_attn.weight" in key or "kv_attn.weight" in key: if (
"c_fc.weight" in key
or "c_proj.weight" in key
or "q_attn.weight" in key
or "kv_attn.weight" in key
):
# Tranpose as we use nn.Linear instead of Conv1D # Tranpose as we use nn.Linear instead of Conv1D
value = value.T value = value.T
@ -99,11 +105,18 @@ class FlashSantacoder(FlashCausalLM):
# Init qkv # Init qkv
if "attn.weight" in final_key: if "attn.weight" in final_key:
module._parameters[param_name] = value.new_empty( module._parameters[param_name] = value.new_empty(
(model.transformer.head_size * (model.transformer.num_heads + 2), value.shape[1]) (
model.transformer.head_size
* (model.transformer.num_heads + 2),
value.shape[1],
)
) )
elif "attn.bias" in final_key: elif "attn.bias" in final_key:
module._parameters[param_name] = value.new_empty( module._parameters[param_name] = value.new_empty(
(model.transformer.head_size * (model.transformer.num_heads + 2)) (
model.transformer.head_size
* (model.transformer.num_heads + 2)
)
) )
# Copy to correct slice # Copy to correct slice
@ -113,11 +126,11 @@ class FlashSantacoder(FlashCausalLM):
module._parameters[param_name][: value.shape[0]] = value module._parameters[param_name][: value.shape[0]] = value
elif "kv_attn.weight" in key: elif "kv_attn.weight" in key:
module._parameters[param_name][ module._parameters[param_name][
model.transformer.head_size * model.transformer.num_heads: model.transformer.head_size * model.transformer.num_heads :
] = value ] = value
elif "kv_attn.bias" in key: elif "kv_attn.bias" in key:
module._parameters[param_name][ module._parameters[param_name][
model.transformer.head_size * model.transformer.num_heads: model.transformer.head_size * model.transformer.num_heads :
] = value ] = value
else: else:
if current_parameter_tensor.shape != value.shape: if current_parameter_tensor.shape != value.shape:

View File

@ -110,6 +110,7 @@ class NextTokenChooser:
class StopSequenceCriteria: class StopSequenceCriteria:
def __init__(self, stop_sequence: str): def __init__(self, stop_sequence: str):
stop_sequence = re.escape(stop_sequence)
self.regex = re.compile(f".*{stop_sequence}$") self.regex = re.compile(f".*{stop_sequence}$")
def __call__(self, output: str) -> bool: def __call__(self, output: str) -> bool: