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 12, 16:00 UTC
[Doc]: fix typos in various files (#29230)
Signed-off-by: Didier Durand <durand.didier@gmail.com> Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Didier Durand committed
Nov 24, 2025 at 12:10 UTC
eca7a8fb59c223c603922be0bd62f5c460972a50
4 files changed
+4
-4
benchmarks/kernels/deepgemm/README.md
+1
-1
@@ -2,7 +2,7 @@
2
3
This directory includes benchmarks between DeepSeek's DeepGEMM block fp8 kernels against vLLM's existing triton and CUTLASS-based kernels.
4
5
-Currently this just includes dense GEMMs and only works on Hopper GPUs.
5
+Currently, this just includes dense GEMMs and only works on Hopper GPUs.
6
7
## Setup
8
vllm/config/vllm.py
+1
-1
@@ -96,7 +96,7 @@ class VllmConfig:
96
"""`torch.compile` and cudagraph capture configuration for the model.
97
98
As a shorthand, one can append compilation arguments via
99
- -0.parameter=arguement such as `-O.mode=3` (same as `-O='{"mode":3}'`).
99
+ -0.parameter=argument such as `-O.mode=3` (same as `-O='{"mode":3}'`).
100
101
You can specify the full compilation config like so:
102
`{"mode": 3, "cudagraph_capture_sizes": [1, 2, 4, 8]}`
vllm/forward_context.py
+1
-1
@@ -153,7 +153,7 @@ class DPMetadata:
153
@contextmanager
154
def sp_local_sizes(self, sequence_parallel_size: int):
155
"""
156
- Context mamager for setting self.local_sizes. Same as self.chunked_sizes
156
+ Context manager for setting self.local_sizes. Same as self.chunked_sizes
157
but without any chunking.
158
"""
159
self.local_sizes = _compute_sp_num_tokens(
vllm/v1/worker/gpu_input_batch.py
+1
-1
@@ -525,7 +525,7 @@ class InputBatch:
525
# NOTE: the following is unsafe
526
# self.token_ids_cpu[i1, ...], self.token_ids_cpu[i2, ...], =\
527
# self.token_ids_cpu[i2, ...], self.token_ids_cpu[i1, ...]
528
- # instead, we need to temporiarily copy the data for one of the indices
528
+ # instead, we need to temporarily copy the data for one of the indices
529
# TODO(lucas): optimize this by only copying valid indices
530
tmp = self.token_ids_cpu[i1, ...].copy()
531
self.token_ids_cpu[i1, ...] = self.token_ids_cpu[i2, ...]