[Doc]: fixing typos in diverse files (#29492)

Signed-off-by: Didier Durand <durand.didier@gmail.com>

vllm/benchmarks/serve.py

@@ -1005,7 +1005,7 @@ def add_cli_args(parser: argparse.ArgumentParser):
         help="Key-value pairs (e.g, --header x-additional-info=0.3.3) "
         "for headers to be passed with each request. These headers override "
         "per backend constants and values set via environment variable, and "
-        "will be overriden by other arguments (such as request ids).",
+        "will be overridden by other arguments (such as request ids).",
     )
     parser.add_argument(
         "--max-concurrency",
@@ -1138,7 +1138,7 @@ def add_cli_args(parser: argparse.ArgumentParser):
         "--percentile-metrics",
         type=str,
         default=None,
-        help="Comma-separated list of selected metrics to report percentils. "
+        help="Comma-separated list of selected metrics to report percentiles. "
         "This argument specifies the metrics to report percentiles. "
         'Allowed metric names are "ttft", "tpot", "itl", "e2el". '
         'If not specified, defaults to "ttft,tpot,itl" for generative models '

vllm/config/parallel.py

@@ -238,9 +238,9 @@ class ParallelConfig:
     cp_kv_cache_interleave_size: int = 1
     """Interleave size of kv_cache storage while using DCP or PCP.
     For `total_cp_rank = pcp_rank * dcp_world_size + dcp_rank`,
-        and `total_cp_world_size = pcp_world_size * dcp_world_szie`.
+        and `total_cp_world_size = pcp_world_size * dcp_world_size`.
     store interleave_size tokens on total_cp_rank i,
-    then store next interleave_size tokens on taotal_cp_rank i+1.
+    then store next interleave_size tokens on total_cp_rank i+1.
     Interleave_size=1: token-level alignment, where token `i` is stored on
         total_cp_rank `i % total_cp_world_size`.
     Interleave_size=block_size: block-level alignment, where tokens are

vllm/lora/punica_wrapper/punica_base.py

@@ -173,7 +173,7 @@ class PunicaWrapperBase(PunicaWrapperABC):
         vocab_size: int,
     ):
         # NOTE We have remove lora extra vocab support for now. So we set
-        # extra_vocab_size alwayzs to 0, and extra_vocab_size will be removed.
+        # extra_vocab_size always to 0, and extra_vocab_size will be removed.
 
         extra_vocab_size = 0
         (

vllm/model_executor/models/adapters.py

@@ -428,7 +428,7 @@ def load_weights_using_from_2_way_softmax(
     )
     if text_config.tie_word_embeddings:
         # embed_tokens is the assumed name for input embeddings. If the model does not
-        # have this attribute, we fallback to get_input_embeddings(), which is used by
+        # have this attribute, we fall back to get_input_embeddings(), which is used by
         # the Transformers modeling backend.
         embed_tokens = (
             model.model.embed_tokens
@@ -486,7 +486,7 @@ def load_weights_no_post_processing(model, weights: Iterable[tuple[str, torch.Te
     )
     if text_config.tie_word_embeddings:
         # embed_tokens is the assumed name for input embeddings. If the model does not
-        # have this attribute, we fallback to get_input_embeddings(), which is used by
+        # have this attribute, we fall back to get_input_embeddings(), which is used by
         # the Transformers modeling backend.
         embed_tokens = (
             model.model.embed_tokens

vllm/v1/sample/tpu/sampler.py

@@ -181,7 +181,7 @@ def apply_top_k_top_p(
     after thresholding the logit using this cut-off, the remaining elements
     shall constitute the top-p set.
 
-    Note: in the case of tie (i.e. multipple cut-off elements present in the
+    Note: in the case of tie (i.e. multiple cut-off elements present in the
     logit), all tie elements are included in the top-p set. In other words,
     this function does not break ties. Instead, these tie tokens have equal
     chance of being chosen during final sampling, so we can consider the tie

vllm/v1/worker/dp_utils.py

@@ -24,12 +24,14 @@ def _get_device_and_group(parallel_config: ParallelConfig):
     device = get_dp_group().device
     group = get_dp_group().device_group
 
-    # Transfering this tensor from GPU to CPU will introduce a GPU sync
+    # Transferring this tensor from GPU to CPU will introduce a GPU sync
     # point that could adversely affect performance of vllm with asynch
     # scheduling. This environment variable exists to quickly disable
     # this optimization if we run into this case.
     if parallel_config.disable_nccl_for_dp_synchronization:
-        logger.info_once("Using CPU all reduce to syncronize DP padding between ranks.")
+        logger.info_once(
+            "Using CPU all reduce to synchronize DP padding between ranks."
+        )
         device = "cpu"
         group = get_dp_group().cpu_group
     return device, group