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 9, 06:15 UTC
[Misc] offest -> offset in comments and variable names (#33444)
Signed-off-by: Russell Bryant <rbryant@redhat.com>
Russell Bryant committed
Jan 30, 2026 at 23:19 UTC
a2ef06e1b3a1c0dc4f7bf6587fbdcf1dad287c31
2 files changed
+3
-3
vllm/v1/attention/backends/mamba2_attn.py
+1
-1
@@ -105,7 +105,7 @@ class Mamba2AttentionMetadata(BaseMambaAttentionMetadata):
105
# Chunk-related metadata (only for prefill)
106
seq_idx_p: torch.Tensor | None = None
107
# cu_chunk_seqlen_p is a tensor of shape (nchunks+1,) that contains, for
108
- # each chunk, its offests into the varlen sequence dimension. It is defined
108
+ # each chunk, its offsets into the varlen sequence dimension. It is defined
109
# such that the i-th chunk contains tokens from cu_chunk_seqlen_p[i] to
110
# cu_chunk_seqlen_p[i+1].
111
cu_chunk_seqlen_p: torch.Tensor | None = None
vllm/v1/spec_decode/eagle.py
+2
-2
@@ -1038,7 +1038,7 @@ class SpecDecodeBaseProposer:
1038
# [0, 1, 2, 3, 4, 5, 6, 7, 8] ->
1039
# [0, 1, 0, 1, 2, 3, 0, 1, 2]
1040
# _r1_ ____r2____ ___r3__
1041
- token_offests = (
1041
+ token_offsets = (
1042
self.token_arange_np[:total_num_tokens] - new_query_start_locs_expanded
1043
)
1044
@@ -1053,7 +1053,7 @@ class SpecDecodeBaseProposer:
1053
# [0, 1, // req 1
1054
# q1 + 0, q1 + 1, q1 + 2, q1 + 3, // req 2
1055
# q1 + q2 + 0, q1 + q2 + 1, q1 + q2 + 2] // req 3
1056
- token_indices_np = token_offests + old_query_start_locs_expanded
1056
+ token_indices_np = token_offsets + old_query_start_locs_expanded
1057
token_indices = torch.from_numpy(token_indices_np).to(device, non_blocking=True)
1058
1059
spec_common_attn_metadata = CommonAttentionMetadata(