A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 04:15 UTC
[Misc] Minor fix in KVCache type (#3652)
Woosuk Kwon committed
Mar 26, 2024 at 23:14 UTC
e66b629c04c38e2743ccac5f14c43bdde016cae1
3 files changed
+4
-8
docs/source/models/adding_model.rst
+2
-2
@@ -56,8 +56,8 @@ Next, you need to rewrite the :code:`forward` methods of your model by following
56
- return_dict: Optional[bool] = None,
57
-) -> Union[Tuple, CausalLMOutputWithPast]:
58
+ positions: torch.Tensor,
59
- + kv_caches: List[KVCache],
60
- + input_metadata: InputMetadata,
59
+ + kv_caches: List[torch.Tensor],
60
+ + attn_metadata: AttentionMetadata,
61
+) -> Optional[SamplerOutput]:
62
63
1. Update the code by considering that :code:`input_ids` and :code:`positions` are now flattened tensors.
vllm/model_executor/models/llava.py
+2
-4
@@ -1,4 +1,4 @@
1
-from typing import List, Optional, Tuple
1
+from typing import List, Optional
2
3
import torch
4
from torch import nn
@@ -19,8 +19,6 @@ from vllm.model_executor.weight_utils import (default_weight_loader,
19
hf_model_weights_iterator)
20
from vllm.sequence import SamplerOutput
21
22
-KVCache = Tuple[torch.Tensor, torch.Tensor]
23
-
22
_KEYS_TO_MODIFY_MAPPING = {
23
"language_model.lm_head": "lm_head",
24
"language_model.model": "language_model",
@@ -102,7 +100,7 @@ class LlavaForConditionalGeneration(nn.Module):
100
self,
101
input_ids: torch.Tensor,
102
positions: torch.Tensor,
105
- kv_caches: List[KVCache],
103
+ kv_caches: List[torch.Tensor],
104
attn_metadata: AttentionMetadata,
105
image_input: Optional[torch.Tensor] = None
106
) -> SamplerOutput: # noqa: E501
vllm/worker/neuron_model_runner.py
-2
@@ -14,8 +14,6 @@ from vllm.utils import (async_tensor_h2d, is_pin_memory_available,
14
15
logger = init_logger(__name__)
16
17
-KVCache = Tuple[torch.Tensor, torch.Tensor]
18
-
17
18
class NeuronModelRunner:
19