mirror of
https://github.com/huggingface/text-generation-inference.git
synced 2025-09-10 03:44:54 +00:00
rename vars
This commit is contained in:
parent
63ddba24b4
commit
f1da19df41
@ -459,8 +459,9 @@ class VlmCausalLMBatch(FlashCausalLMBatch):
|
||||
image_positions = []
|
||||
num_images = len(image_texts)
|
||||
|
||||
input_ids_t = torch.as_tensor(input_ids, dtype=torch.int32)
|
||||
input_ids_t = torch.as_tensor(input_ids)
|
||||
img_start_token_pos = torch.where(input_ids_t.eq(img_start_token))[0]
|
||||
num_tokens = input_ids_t.numel()
|
||||
|
||||
last_pos = 0
|
||||
for i in range(num_images):
|
||||
@ -470,25 +471,25 @@ class VlmCausalLMBatch(FlashCausalLMBatch):
|
||||
if config.model_type == "gemma3":
|
||||
img_text = img_text.replace("\n\n", "")
|
||||
|
||||
tokens = tokenizer(img_text, add_special_tokens=False)["input_ids"]
|
||||
length = len(tokens)
|
||||
tokens = tokenizer(img_text, add_special_tokens=False, return_tensors="pt")[
|
||||
"input_ids"
|
||||
][0]
|
||||
length = tokens.numel()
|
||||
|
||||
assert length <= len(
|
||||
input_ids
|
||||
), f"{length} > {len(input_ids)} Image is truncated, try increasing --max-batch-prefill-tokens"
|
||||
assert (
|
||||
length <= num_tokens
|
||||
), f"{length} > {num_tokens} Image is truncated, try increasing --max-batch-prefill-tokens"
|
||||
|
||||
pos = torch.searchsorted(img_start_token_pos, last_pos, right=False)
|
||||
index = img_start_token_pos[pos]
|
||||
|
||||
assert (
|
||||
input_ids[index : index + length] == tokens
|
||||
assert torch.equal(
|
||||
input_ids_t[index : index + length], tokens
|
||||
), "Image tokens not found in input_ids"
|
||||
|
||||
num_placeholder_tokens = tokens.count(config.image_token_index)
|
||||
is_embed = tokens == config.image_token_index
|
||||
num_placeholder_tokens = int(is_embed.sum())
|
||||
if num_placeholder_tokens == length:
|
||||
is_embed = None
|
||||
else:
|
||||
is_embed = torch.as_tensor(tokens) == config.image_token_index
|
||||
|
||||
pos = ImagePositions(
|
||||
offset=index,
|
||||
|
Loading…
Reference in New Issue
Block a user