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 15, 09:45 UTC
fix: typos (#18151)
Signed-off-by: omahs <73983677+omahs@users.noreply.github.com>
omahs committed
May 15, 2025 at 11:16 UTC
a9944aabfa0eb0f133cf869b3ed5defb44ed7d33
10 files changed
+18
-18
csrc/attention/attention_kernels.cuh
+2
-2
@@ -172,7 +172,7 @@ __device__ void paged_attention_kernel(
172
173
// Load the query to registers.
174
// Each thread in a thread group has a different part of the query.
175
- // For example, if the the thread group size is 4, then the first thread in
175
+ // For example, if the thread group size is 4, then the first thread in
176
// the group has 0, 4, 8, ... th vectors of the query, and the second thread
177
// has 1, 5, 9, ... th vectors of the query, and so on. NOTE(woosuk): Because
178
// q is split from a qkv tensor, it may not be contiguous.
@@ -259,7 +259,7 @@ __device__ void paged_attention_kernel(
259
260
// Load a key to registers.
261
// Each thread in a thread group has a different part of the key.
262
- // For example, if the the thread group size is 4, then the first thread in
262
+ // For example, if the thread group size is 4, then the first thread in
263
// the group has 0, 4, 8, ... th vectors of the key, and the second thread
264
// has 1, 5, 9, ... th vectors of the key, and so on.
265
for (int i = 0; i < NUM_TOKENS_PER_THREAD_GROUP; i++) {
examples/offline_inference/chat_with_tools.py
+2
-2
@@ -68,7 +68,7 @@ def get_current_weather(city: str, state: str, unit: 'str'):
68
"partly cloudly, with highs in the 90's.")
69
70
71
-tool_funtions = {"get_current_weather": get_current_weather}
71
+tool_functions = {"get_current_weather": get_current_weather}
72
73
tools = [{
74
"type": "function",
@@ -122,7 +122,7 @@ messages.append({
122
# above defined function
123
tool_calls = json.loads(output)
124
tool_answers = [
125
- tool_funtions[call['name']](**call['arguments']) for call in tool_calls
125
+ tool_functions[call['name']](**call['arguments']) for call in tool_calls
126
]
127
128
# append the answer as a tool message and let the LLM give you an answer
tests/lora/test_lora_huggingface.py
+1
-1
@@ -30,7 +30,7 @@ def test_load_checkpoints_from_huggingface(lora_fixture_name, request):
30
31
lora_path = get_adapter_absolute_path(lora_name)
32
33
- # lora loading should work for either absolute path and hugggingface id.
33
+ # lora loading should work for either absolute path and huggingface id.
34
peft_helper = PEFTHelper.from_local_dir(lora_path, 4096)
35
lora_model = LoRAModel.from_local_checkpoint(
36
lora_path,
tests/model_executor/weight_utils.py
+3
-3
@@ -20,11 +20,11 @@ def test_hf_transfer_auto_activation():
20
try:
21
# enable hf hub transfer if available
22
import hf_transfer # type: ignore # noqa
23
- HF_TRANFER_ACTIVE = True
23
+ HF_TRANSFER_ACTIVE = True
24
except ImportError:
25
- HF_TRANFER_ACTIVE = False
25
+ HF_TRANSFER_ACTIVE = False
26
assert (huggingface_hub.constants.HF_HUB_ENABLE_HF_TRANSFER ==
27
- HF_TRANFER_ACTIVE)
27
+ HF_TRANSFER_ACTIVE)
28
29
30
def test_download_weights_from_hf():
vllm/config.py
+1
-1
@@ -297,7 +297,7 @@ class ModelConfig:
297
- 1K -> 1024\n
298
- 25.6k -> 25,600"""
299
spec_target_max_model_len: Optional[int] = None
300
- """Specify the the maximum length for spec decoding draft models."""
300
+ """Specify the maximum length for spec decoding draft models."""
301
quantization: Optional[QuantizationMethods] = None
302
"""Method used to quantize the weights. If `None`, we first check the
303
`quantization_config` attribute in the model config file. If that is
vllm/lora/ops/triton_ops/lora_expand_op.py
+1
-1
@@ -153,7 +153,7 @@ def _lora_expand(
153
lora_token_start_loc (torch.Tensor): A cumulative sum of
154
num_tokens_per_lora. lora_token_start_loc[0] is always 0 so that
155
lora_token_start_loc[i], along with num_tokens_per_lora[i]
156
- identifies the the region in token_indices_sorted_by_lora_ids that
156
+ identifies the region in token_indices_sorted_by_lora_ids that
157
LoRA lora_ids[i] should process.
158
lora_ids (torch.Tensor): LoRA ids to process.
159
no_lora_flag_cpu (torch.Tensor): A CPU tensor of size 1, that indicates
vllm/model_executor/layers/mamba/mamba_mixer2.py
+1
-1
@@ -142,7 +142,7 @@ def mamba_v2_sharded_weight_loader(
142
) -> LoaderFunction:
143
"""Create a weight loader for mamba v2. This ensures that the projections
144
are correctly sharded so that they can be split into x, B, C. It also
145
- ensures the the all the groups corresponding to a head shard is placed
145
+ ensures that all the groups corresponding to a head shard is placed
146
together with it.
147
"""
148
vllm/model_executor/models/granite_speech.py
+2
-2
@@ -21,7 +21,7 @@
21
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
22
# See the License for the specific language governing permissions and
23
# limitations under the License.
24
-"""Inference-only IBM Granite speeech model."""
24
+"""Inference-only IBM Granite speech model."""
25
import math
26
from collections.abc import Iterable, Mapping
27
from typing import Optional, TypedDict, Union
@@ -626,7 +626,7 @@ class GraniteSpeechForConditionalGeneration(
626
audio_embed_sizes: torch.Tensor,
627
) -> torch.Tensor:
628
"""Calculate the input features mask, which will generally be used
629
- to mask the the padded features for all entries in the batch except
629
+ to mask the padded features for all entries in the batch except
630
for those with the most audio features.
631
632
Args:
vllm/model_executor/models/phi4mm_audio.py
+4
-4
@@ -91,9 +91,9 @@ class ConformerEncoderLayer(nn.Module):
91
if set to True, use GLULinear module,
92
otherwise, used GLUPointWiseConv module.
93
default to False.
94
- attention_innner_dim: int, optional
94
+ attention_inner_dim: int, optional
95
if equal to -1, attention dim for linears k/q/v is
96
- equal to d_model. otherwise attention_innner_dim is used.
96
+ equal to d_model. otherwise attention_inner_dim is used.
97
default -1.
98
attention_glu_type: str, optional
99
activation function for glu used in the multihead attention,
@@ -148,7 +148,7 @@ class ConformerEncoderLayer(nn.Module):
148
conv_glu_type="sigmoid",
149
bias_in_glu=True,
150
linear_glu_in_convm=False,
151
- attention_innner_dim=-1,
151
+ attention_inner_dim=-1,
152
attention_glu_type="swish",
153
activation_checkpointing="",
154
export=False,
@@ -169,7 +169,7 @@ class ConformerEncoderLayer(nn.Module):
169
n_head,
170
d_model,
171
dropout_rate,
172
- attention_innner_dim,
172
+ attention_inner_dim,
173
attention_glu_type,
174
bias_in_glu,
175
use_pt_scaled_dot_product_attention=
vllm/v1/request.py
+1
-1
@@ -72,7 +72,7 @@ class Request:
72
assert len(self.mm_inputs) == len(self.mm_hashes)
73
74
# Read-only views
75
- # Prevent directly appending to the these lists since
75
+ # Prevent directly appending to these lists since
76
# they should also be updated simultaneously.
77
self.output_token_ids = ConstantList(self._output_token_ids)
78
self.all_token_ids = ConstantList(self._all_token_ids)