xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-05-22 08:04:27 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix some typos (#24616)

Browse Source 
Signed-off-by: co63oc <co63oc@users.noreply.github.com>
This commit is contained in:
co63oc 2025-09-12 01:48:46 +08:00 committed by GitHub
parent c1eda615ba
commit e26fef8397
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tests/compile/test_basic_correctness.py
View File

@ -23,7 +23,7 @@ class TestSetting:
fullgraph: bool fullgraph: bool
# we cannot afford testing the full Catesian product # we cannot afford testing the full Cartesian product
# of all models and all levels # of all models and all levels
@pytest.mark.parametrize( @pytest.mark.parametrize(
"test_setting", "test_setting",

2
tests/kernels/mamba/test_mamba_ssm_ssd.py
View File

@ -345,7 +345,7 @@ def test_mamba_chunk_scan_cont_batch_prefill_chunking(chunk_size, seqlens):
# in the mamba2 ssd kernels, by comparing concatenation (in the sequence # in the mamba2 ssd kernels, by comparing concatenation (in the sequence
# dimension) of chunked results with the full sequence result. # dimension) of chunked results with the full sequence result.
# It is different from test_mamba_chunk_scan_cont_batch by: # It is different from test_mamba_chunk_scan_cont_batch by:
# 1. Not using the naive torch implementaion (ssd_minimal_discrete) to get # 1. Not using the naive torch implementation (ssd_minimal_discrete) to get
# reference outputs. Instead, it compares chunked kernel outputs to full # reference outputs. Instead, it compares chunked kernel outputs to full
# sequence kernel outputs. This is the most straightforward way to # sequence kernel outputs. This is the most straightforward way to
# assert chunked prefill correctness. # assert chunked prefill correctness.

2
vllm/model_executor/layers/fla/ops/cumsum.py
View File

@ -179,7 +179,7 @@ def chunk_local_cumsum_vector(
def grid(meta): def grid(meta):
return (triton.cdiv(meta['S'], meta['BS']), NT, B * H) return (triton.cdiv(meta['S'], meta['BS']), NT, B * H)
# keep cummulative normalizer in fp32 # keep cumulative normalizer in fp32
# this kernel is equivalent to # this kernel is equivalent to
# g = g.view(B, H, NT, BT, -1).cumsum(-2).view(B, H, T, -1) # g = g.view(B, H, NT, BT, -1).cumsum(-2).view(B, H, T, -1)
chunk_local_cumsum_vector_kernel[grid](g_org, chunk_local_cumsum_vector_kernel[grid](g_org,

2
vllm/v1/attention/backends/mla/common.py
View File

@ -1322,7 +1322,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
k_scale: torch.Tensor, k_scale: torch.Tensor,
dcp_world_size: int, dcp_world_size: int,
): ):
assert k_scale is None, "DCP not support sacled kvcache now." assert k_scale is None, "DCP not support scaled kvcache now."
assert attn_metadata.prefill is not None assert attn_metadata.prefill is not None
prefill_metadata = attn_metadata.prefill prefill_metadata = attn_metadata.prefill
assert prefill_metadata.chunked_context is not None assert prefill_metadata.chunked_context is not None

4
vllm/v1/worker/block_table.py
View File

@ -112,9 +112,9 @@ class BlockTable:
# tokens. # tokens.
virtual_block_offsets = positions % virtual_block_size virtual_block_offsets = positions % virtual_block_size
mask = virtual_block_offsets % self.dcp_world_size == self.dcp_rank mask = virtual_block_offsets % self.dcp_world_size == self.dcp_rank
# Calcuate local block_offsets # Calculate local block_offsets
block_offsets = virtual_block_offsets // self.dcp_world_size block_offsets = virtual_block_offsets // self.dcp_world_size
# Calcuate slot_mapping # Calculate slot_mapping
slot_mapping = block_numbers * self.block_size + block_offsets slot_mapping = block_numbers * self.block_size + block_offsets
# Write final slots, use -1 for not-local # Write final slots, use -1 for not-local
self.slot_mapping_np[:req_indices.shape[0]] = np.where( self.slot_mapping_np[:req_indices.shape[0]] = np.where(