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 14, 05:30 UTC
fix some typos (#24616)
Signed-off-by: co63oc <co63oc@users.noreply.github.com>
co63oc committed
Sep 12, 2025 at 01:48 UTC
e26fef83977d0361522dd294c8e9659e57b2eb5f
5 files changed
+6
-6
tests/compile/test_basic_correctness.py
+1
-1
@@ -23,7 +23,7 @@ class TestSetting:
23
fullgraph: bool
24
25
26
-# we cannot afford testing the full Catesian product
26
+# we cannot afford testing the full Cartesian product
27
# of all models and all levels
28
@pytest.mark.parametrize(
29
"test_setting",
tests/kernels/mamba/test_mamba_ssm_ssd.py
+1
-1
@@ -345,7 +345,7 @@ def test_mamba_chunk_scan_cont_batch_prefill_chunking(chunk_size, seqlens):
345
# in the mamba2 ssd kernels, by comparing concatenation (in the sequence
346
# dimension) of chunked results with the full sequence result.
347
# It is different from test_mamba_chunk_scan_cont_batch by:
348
- # 1. Not using the naive torch implementaion (ssd_minimal_discrete) to get
348
+ # 1. Not using the naive torch implementation (ssd_minimal_discrete) to get
349
# reference outputs. Instead, it compares chunked kernel outputs to full
350
# sequence kernel outputs. This is the most straightforward way to
351
# assert chunked prefill correctness.
vllm/model_executor/layers/fla/ops/cumsum.py
+1
-1
@@ -179,7 +179,7 @@ def chunk_local_cumsum_vector(
179
def grid(meta):
180
return (triton.cdiv(meta['S'], meta['BS']), NT, B * H)
181
182
- # keep cummulative normalizer in fp32
182
+ # keep cumulative normalizer in fp32
183
# this kernel is equivalent to
184
# g = g.view(B, H, NT, BT, -1).cumsum(-2).view(B, H, T, -1)
185
chunk_local_cumsum_vector_kernel[grid](g_org,
vllm/v1/attention/backends/mla/common.py
+1
-1
@@ -1322,7 +1322,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
1322
k_scale: torch.Tensor,
1323
dcp_world_size: int,
1324
):
1325
- assert k_scale is None, "DCP not support sacled kvcache now."
1325
+ assert k_scale is None, "DCP not support scaled kvcache now."
1326
assert attn_metadata.prefill is not None
1327
prefill_metadata = attn_metadata.prefill
1328
assert prefill_metadata.chunked_context is not None
vllm/v1/worker/block_table.py
+2
-2
@@ -112,9 +112,9 @@ class BlockTable:
112
# tokens.
113
virtual_block_offsets = positions % virtual_block_size
114
mask = virtual_block_offsets % self.dcp_world_size == self.dcp_rank
115
- # Calcuate local block_offsets
115
+ # Calculate local block_offsets
116
block_offsets = virtual_block_offsets // self.dcp_world_size
117
- # Calcuate slot_mapping
117
+ # Calculate slot_mapping
118
slot_mapping = block_numbers * self.block_size + block_offsets
119
# Write final slots, use -1 for not-local
120
self.slot_mapping_np[:req_indices.shape[0]] = np.where(