Qwen2.5-VL eagle3 infer (#8801)
This commit is contained in:
@@ -518,6 +518,9 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
|
||||
self.logits_processor = LogitsProcessor(config)
|
||||
self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
|
||||
|
||||
# For EAGLE3 support
|
||||
self.capture_aux_hidden_states = False
|
||||
|
||||
def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
|
||||
pattern = MultiModalityDataPaddingPatternMultimodalTokens()
|
||||
return pattern.pad_input_tokens(input_ids, mm_inputs)
|
||||
@@ -588,9 +591,13 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
|
||||
positions=positions,
|
||||
)
|
||||
|
||||
aux_hidden_states = None
|
||||
if self.capture_aux_hidden_states:
|
||||
hidden_states, aux_hidden_states = hidden_states
|
||||
|
||||
if not get_embedding:
|
||||
return self.logits_processor(
|
||||
input_ids, hidden_states, self.lm_head, forward_batch
|
||||
input_ids, hidden_states, self.lm_head, forward_batch, aux_hidden_states
|
||||
)
|
||||
else:
|
||||
return self.pooler(hidden_states, forward_batch)
|
||||
@@ -644,5 +651,21 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
|
||||
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||
weight_loader(param, loaded_weight)
|
||||
|
||||
def get_embed_and_head(self):
|
||||
return self.model.embed_tokens.weight, self.lm_head.weight
|
||||
|
||||
def set_eagle3_layers_to_capture(self, layer_ids: Optional[List[int]] = None):
|
||||
self.capture_aux_hidden_states = True
|
||||
self.model.capture_aux_hidden_states = True
|
||||
if layer_ids is None:
|
||||
num_layers = self.config.num_hidden_layers
|
||||
self.model.layers_to_capture = [
|
||||
2,
|
||||
num_layers // 2,
|
||||
num_layers - 3,
|
||||
] # Specific layers for EAGLE3 support
|
||||
else:
|
||||
self.model.layers_to_capture = [val + 1 for val in layer_ids]
|
||||
|
||||
|
||||
EntryClass = [Qwen2_5_VLForConditionalGeneration]
|
||||
|
||||
Reference in New Issue
Block a user