Fixing qwen2.5 VL.

This commit is contained in:
Nicolas Patry 2025-03-11 10:57:58 +01:00
parent 094975c3a8
commit 911910c5d2
No known key found for this signature in database
GPG Key ID: 4242CEF24CB6DBF9

View File

@ -313,7 +313,6 @@ class Qwen2_5_VLVisionConfig(PretrainedConfig):
class Qwen2_5_VLConfig(PretrainedConfig): class Qwen2_5_VLConfig(PretrainedConfig):
def __init__( def __init__(
self, self,
vocab_size=152064, vocab_size=152064,
@ -702,7 +701,6 @@ class Qwen2_5VisionModel(nn.Module):
pixel_values: torch.Tensor, pixel_values: torch.Tensor,
grid_thw: Optional[torch.LongTensor] = None, grid_thw: Optional[torch.LongTensor] = None,
) -> torch.Tensor: ) -> torch.Tensor:
# reshape the input tensor for processing # reshape the input tensor for processing
shape = ( shape = (
-1, -1,
@ -967,6 +965,7 @@ class Qwen2_5VLForConditionalGeneration(nn.Module):
max_s=max_s, max_s=max_s,
true_max_s=max_s, true_max_s=max_s,
prefill_cache_indices=prefill_cache_indices, prefill_cache_indices=prefill_cache_indices,
adapter_data=adapter_data,
) )
if lm_head_indices is not None: if lm_head_indices is not None:
hidden_states = hidden_states[lm_head_indices] hidden_states = hidden_states[lm_head_indices]