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 8, 17:45 UTC
Fix spelling errors (#33978)
lukec committed
Feb 7, 2026 at 15:58 UTC
15a0b9e570dc8bfe716a7d76d50716898123dbae
5 files changed
+10
-10
tests/kernels/moe/test_cutedsl_moe.py
+1
-1
@@ -139,7 +139,7 @@ def prepare_inputs(
139
masked_m.append(mask.sum())
140
141
masked_m = torch.tensor(masked_m, dtype=torch.int32)
142
- # Intialize the hidden_states_3d with ones instead of empty to avoid nan
142
+ # Initialize the hidden_states_3d with ones instead of empty to avoid nan
143
# issue.
144
hidden_states_3d = torch.ones(
145
(num_experts, max(masked_m), hidden_states.shape[1]), dtype=hidden_states.dtype
tests/kernels/moe/test_moe_align_block_size.py
+1
-1
@@ -334,7 +334,7 @@ def test_batched_moe_align_block_size(
334
ref_expert_ids = torch.empty((Msum // block_size,), dtype=torch.int32)
335
ref_num_tokens_post_pad = torch.empty((1,), dtype=torch.int32)
336
337
- # Intialize
337
+ # Initialize
338
sentinel = E * max_tokens_per_batch
339
ref_sorted_ids.fill_(sentinel)
340
ref_expert_ids.fill_(-1)
tests/reasoning/test_hunyuan_reasoning_parser.py
+3
-3
@@ -12,7 +12,7 @@ START_REASONING = "<think>\n"
12
START_RESPONSE = "\n</think>\n<answer>\n"
13
END_RESPONSE = "\n</answer>"
14
15
-NO_REASONING_QUICK_THROUGHT = {
15
+NO_REASONING_QUICK_THOUGHT = {
16
"output": f"{START_REASONING}{START_RESPONSE}This is the rest{END_RESPONSE}", # noqa: E501
17
"reasoning": None,
18
"content": "This is the rest",
@@ -81,7 +81,7 @@ TEST_CASES = [
81
NO_REASONING,
82
id="no_reasoning",
83
),
84
- pytest.param(False, NO_REASONING_QUICK_THROUGHT, id="no_reasoning_quick"),
84
+ pytest.param(False, NO_REASONING_QUICK_THOUGHT, id="no_reasoning_quick"),
85
pytest.param(
86
False,
87
MULTIPLE_LINES,
@@ -117,7 +117,7 @@ TEST_CASES = [
117
NO_REASONING,
118
id="no_reasoning_streaming",
119
),
120
- pytest.param(True, NO_REASONING_QUICK_THROUGHT, id="no_reasoning_quick_stream"),
120
+ pytest.param(True, NO_REASONING_QUICK_THOUGHT, id="no_reasoning_quick_stream"),
121
pytest.param(
122
True,
123
MULTIPLE_LINES,
vllm/model_executor/layers/fused_moe/oracle/fp8.py
+1
-1
@@ -333,7 +333,7 @@ def select_fp8_moe_backend(
333
334
# TODO(rob): per discussion with TPU team, we need a way to register
335
# MoE backends by OOT plugins, rather than having an explicit list
336
- # of AVAILBLE_BACKENDS. Enabling returning `Fp8MoeBackend.NONE` is
336
+ # of AVAILABLE_BACKENDS. Enabling returning `Fp8MoeBackend.NONE` is
337
# a temporary measure until these register APIs are complete.
338
if current_platform.is_cuda() or current_platform.is_rocm():
339
raise NotImplementedError(
vllm/v1/attention/ops/flashmla.py
+4
-4
@@ -52,8 +52,8 @@ def is_flashmla_dense_supported() -> tuple[bool, str | None]:
52
"""
53
Return: is_supported_flag, unsupported_reason (optional).
54
"""
55
- is_availble, maybe_reason = _is_flashmla_available()
56
- if not is_availble:
55
+ is_available, maybe_reason = _is_flashmla_available()
56
+ if not is_available:
57
return False, maybe_reason
58
if not current_platform.is_device_capability_family(90):
59
return False, "FlashMLA Dense is only supported on Hopper devices."
@@ -64,8 +64,8 @@ def is_flashmla_sparse_supported() -> tuple[bool, str | None]:
64
"""
65
Return: is_supported_flag, unsupported_reason (optional).
66
"""
67
- is_availble, maybe_reason = _is_flashmla_available()
68
- if not is_availble:
67
+ is_available, maybe_reason = _is_flashmla_available()
68
+ if not is_available:
69
return False, maybe_reason
70
if not (
71
current_platform.is_device_capability_family(90)