Gameur/Beta
1
0
Fork 0
Code Issues 1 Pull requests Projects Releases Packages Wiki Activity Actions

Voice et bot modif

Browse source
This commit is contained in:
pi 2026-06-16 17:09:34 +00:00
parent 189d56026b
commit 7333a22bcd
10774 changed files with 634644 additions and 933308 deletions
Download patch file Download diff file Expand all files Collapse all files

19
venv/lib/python3.12/site-packages/ctranslate2/specs/transformer_spec.py
View file

@ -101,6 +101,7 @@ class TransformerEncoderSpec(model_spec.LayerSpec):
rms_norm=rms_norm,
num_heads_kv=num_heads_kv,
head_dim=head_dim,
sliding_window=sliding_window,
rotary_dim=rotary_dim,
rotary_interleave=rotary_interleave,
rotary_scaling_type=rotary_scaling_type,
@ -151,7 +152,9 @@ class TransformerDecoderSpec(model_spec.LayerSpec):
quant_group_size: Optional[int] = None,
quant_bits: Optional[int] = None,
qk_norm: bool = False,
v_norm: bool = False,
external_pre_post_encoder_layers: Optional[bool] = False,
merged_encoder_attention: bool = False,
):
"""Initializes a Transformer decoder specification.
@ -262,11 +265,15 @@ class TransformerDecoderSpec(model_spec.LayerSpec):
head_dim=head_dim,
sliding_window=sliding_window,
qk_norm=qk_norm,
v_norm=v_norm,
external_pre_post_encoder_layers=external_pre_post_encoder_layers,
merged_encoder_attention=merged_encoder_attention,
)
for _ in range(num_layers)
]
self.start_from_zero_embedding = False
self.final_logit_softcapping = model_spec.OPTIONAL
self._config["multi_query_attention"] = multi_query_attention or (
num_heads_kv != num_heads
)
@ -275,7 +282,7 @@ class TransformerDecoderSpec(model_spec.LayerSpec):
self.project_in = common_spec.LinearSpec()
self.project_out = common_spec.LinearSpec()
if quant_type is not None:
if quant_type:
self._config["quantization_type"] = quant_type
self._config["quantization_bits"] = quant_bits
self._config["quantization_group_size"] = quant_group_size
@ -358,7 +365,9 @@ class TransformerDecoderLayerSpec(model_spec.LayerSpec):
head_dim=None,
sliding_window=None,
qk_norm=False,
v_norm=False,
external_pre_post_encoder_layers=False,
merged_encoder_attention=False,
):
self.self_attention = attention_spec.MultiHeadAttentionSpec(
self_attention=True,
@ -376,9 +385,11 @@ class TransformerDecoderLayerSpec(model_spec.LayerSpec):
head_dim=head_dim,
sliding_window=sliding_window,
qk_norm=qk_norm,
v_norm=v_norm,
merged_encoder_attention=merged_encoder_attention,
)
if with_encoder_attention:
if with_encoder_attention and not merged_encoder_attention:
self.attention = attention_spec.MultiHeadAttentionSpec(
rms_norm=rms_norm,
num_heads_kv=num_heads_kv,
@ -426,6 +437,8 @@ class TransformerDecoderLayerSpec(model_spec.LayerSpec):
delattr(self.self_attention, "layer_norm")
delattr(self.ffn, "layer_norm")
self.layer_scalar = model_spec.OPTIONAL
class FeedForwardSpec(model_spec.LayerSpec):
def __init__(self, glu=False, rms_norm=False):
@ -646,6 +659,7 @@ class TransformerDecoderModelSpec(model_spec.LanguageModelSpec):
quant_group_size: Optional[int] = None,
quant_bits: Optional[int] = None,
qk_norm: bool = False,
v_norm: bool = False,
):
"""Creates a Transformer decoder model specification.
@ -721,6 +735,7 @@ class TransformerDecoderModelSpec(model_spec.LanguageModelSpec):
quant_group_size=quant_group_size,
quant_bits=quant_bits,
qk_norm=qk_norm,
v_norm=v_norm,
)
return cls(decoder)