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 17, 04:15 UTC
chore(vllm): codespell for spell checking (#2820)
Massimiliano Pronesti committed
Feb 22, 2024 at 02:56 UTC
93dc5a287086299a124e9f1f6fac75458ae0acbd
16 files changed
+85
-25
.github/workflows/ruff.yml
+4
-1
@@ -25,7 +25,10 @@ jobs:
25
- name: Install dependencies
26
run: |
27
python -m pip install --upgrade pip
28
- pip install ruff==0.1.5
28
+ pip install ruff==0.1.5 codespell==2.2.6 tomli==2.0.1
29
- name: Analysing the code with ruff
30
run: |
31
ruff vllm tests
32
+ - name: Spelling check with codespell
33
+ run: |
34
+ codespell --toml pyproject.toml
\ No newline at end of file
benchmarks/benchmark_serving.py
+1
-1
@@ -375,7 +375,7 @@ if __name__ == "__main__":
375
parser.add_argument(
376
"--disable-tqdm",
377
action="store_true",
378
- help="Specify to disbale tqdm progress bar.",
378
+ help="Specify to disable tqdm progress bar.",
379
)
380
parser.add_argument(
381
"--save-result",
format.sh
+48
-3
@@ -24,6 +24,7 @@ builtin cd "$ROOT" || exit 1
24
YAPF_VERSION=$(yapf --version | awk '{print $2}')
25
RUFF_VERSION=$(ruff --version | awk '{print $2}')
26
MYPY_VERSION=$(mypy --version | awk '{print $2}')
27
+CODESPELL_VERSION=$(codespell --version)
28
29
# # params: tool name, tool version, required version
30
tool_version_check() {
@@ -36,6 +37,7 @@ tool_version_check() {
37
tool_version_check "yapf" $YAPF_VERSION "$(grep yapf requirements-dev.txt | cut -d'=' -f3)"
38
tool_version_check "ruff" $RUFF_VERSION "$(grep "ruff==" requirements-dev.txt | cut -d'=' -f3)"
39
tool_version_check "mypy" "$MYPY_VERSION" "$(grep mypy requirements-dev.txt | cut -d'=' -f3)"
40
+tool_version_check "codespell" "$CODESPELL_VERSION" "$(grep codespell requirements-dev.txt | cut -d'=' -f3)"
41
42
YAPF_FLAGS=(
43
'--recursive'
@@ -93,6 +95,47 @@ echo 'vLLM yapf: Done'
95
# echo 'vLLM mypy:'
96
# mypy
97
98
+# check spelling of specified files
99
+spell_check() {
100
+ codespell "$@"
101
+}
102
+
103
+spell_check_all(){
104
+ codespell --toml pyproject.toml
105
+}
106
+
107
+# Spelling check of files that differ from main branch.
108
+spell_check_changed() {
109
+ # The `if` guard ensures that the list of filenames is not empty, which
110
+ # could cause ruff to receive 0 positional arguments, making it hang
111
+ # waiting for STDIN.
112
+ #
113
+ # `diff-filter=ACM` and $MERGEBASE is to ensure we only lint files that
114
+ # exist on both branches.
115
+ MERGEBASE="$(git merge-base origin/main HEAD)"
116
+
117
+ if ! git diff --diff-filter=ACM --quiet --exit-code "$MERGEBASE" -- '*.py' '*.pyi' &>/dev/null; then
118
+ git diff --name-only --diff-filter=ACM "$MERGEBASE" -- '*.py' '*.pyi' | xargs \
119
+ codespell
120
+ fi
121
+}
122
+
123
+# Run Codespell
124
+## This flag runs spell check of individual files. --files *must* be the first command line
125
+## arg to use this option.
126
+if [[ "$1" == '--files' ]]; then
127
+ spell_check "${@:2}"
128
+ # If `--all` is passed, then any further arguments are ignored and the
129
+ # entire python directory is linted.
130
+elif [[ "$1" == '--all' ]]; then
131
+ spell_check_all
132
+else
133
+ # Check spelling only of the files that changed in last commit.
134
+ spell_check_changed
135
+fi
136
+echo 'vLLM codespell: Done'
137
+
138
+
139
# Lint specified files
140
lint() {
141
ruff "$@"
@@ -117,9 +160,9 @@ lint_changed() {
160
}
161
162
# Run Ruff
120
-echo 'vLLM Ruff:'
121
-## This flag lints individual files. --files *must* be the first command line
122
-## arg to use this option.
163
+echo 'vLLM ruff:'
164
+### This flag lints individual files. --files *must* be the first command line
165
+### arg to use this option.
166
if [[ "$1" == '--files' ]]; then
167
lint "${@:2}"
168
# If `--all` is passed, then any further arguments are ignored and the
@@ -139,3 +182,5 @@ if ! git diff --quiet &>/dev/null; then
182
183
exit 1
184
fi
185
+
186
+
mypy.ini
deleted
-8
@@ -1,8 +0,0 @@
1
-[mypy]
2
-python_version = 3.8
3
-
4
-ignore_missing_imports = True
5
-
6
-files = vllm
7
-# TODO(woosuk): Include the code from Megatron and HuggingFace.
8
-exclude = vllm/model_executor/parallel_utils/|vllm/model_executor/models/
pyproject.toml
+18
@@ -31,4 +31,22 @@ ignore = [
31
"E731",
32
# line too long, handled by black formatting
33
"E501",
34
+ # .strip() with multi-character strings
35
+ "B005",
36
+ # Loop control variable not used within loop body
37
+ "B007",
38
]
39
+
40
+[tool.mypy]
41
+python_version = "3.8"
42
+
43
+ignore_missing_imports = true
44
+
45
+files = "vllm"
46
+# TODO(woosuk): Include the code from Megatron and HuggingFace.
47
+exclude = "vllm/model_executor/parallel_utils/|vllm/model_executor/models/"
48
+
49
+
50
+[tool.codespell]
51
+ignore-words-list = "dout, te, indicies"
52
+skip = "./tests/prompts"
requirements-dev.txt
+2
@@ -1,7 +1,9 @@
1
# formatting
2
yapf==0.32.0
3
toml==0.10.2
4
+tomli==2.0.1
5
ruff==0.1.5
6
+codespell==2.2.6
7
8
# type checking
9
mypy==0.991
tests/lora/test_layers.py
+1
-1
@@ -279,7 +279,7 @@ def test_embeddings_with_new_embeddings(dist_init, num_loras, device) -> None:
279
256,
280
org_num_embeddings=512)
281
expanded_embedding.weight.data[:512, :] = embedding_data
282
- # We need to deepcopy the embedding as it will be modifed
282
+ # We need to deepcopy the embedding as it will be modified
283
# in place
284
lora_embedding = VocabParallelEmbeddingWithLoRA(
285
deepcopy(expanded_embedding))
tests/lora/test_llama.py
+2
-2
@@ -15,7 +15,7 @@ def do_sample(llm, lora_path: str, lora_id: int):
15
"[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_95 (one_mora VARCHAR, gloss VARCHAR, accented_mora VARCHAR)\n\n question: What is the one mora for a low tone mora with a gloss of /˩okiru/ [òkìɽɯ́]? [/user] [assistant]",
16
"[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE candidate (people_id VARCHAR, unsure_rate INTEGER); CREATE TABLE people (sex VARCHAR, people_id VARCHAR)\n\n question: which gender got the highest average uncertain ratio. [/user] [assistant]",
17
"[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_60 (pick INTEGER, former_wnba_team VARCHAR)\n\n question: What pick was a player that previously played for the Minnesota Lynx? [/user] [assistant]",
18
- "[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the womens doubles for werner schlager [/user] [assistant]"
18
+ "[user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the women's doubles for werner schlager [/user] [assistant]"
19
]
20
sampling_params = vllm.SamplingParams(temperature=0,
21
max_tokens=256,
@@ -53,7 +53,7 @@ def test_llama_lora(sql_lora_files, tp_size):
53
"\n\n answer: 1\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_96 (one_mora VARCHAR, gloss VARCHAR, accented_mora VARCHAR)\n\n question: What is the one mora for a high tone mora with a gloss of /˧kot/ [kòt]? [/user] [assistant]\n\n answer: 2\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_97 (one_mora VARCHAR, gloss VARCHAR, accented_mora VARCHAR)\n\n question: What is the one mora for a high tone mora with a gloss of /˧kot/ [kòt]? [/user] [assistant]\n\n answer: 2\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_98 (one_mora VARCHAR, gloss VARCHAR, accented_mora VARCHAR)\n\n question: What is the one m",
54
" Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE candidate (people_id VARCHAR, unsure_rate INTEGER); CREATE TABLE people (sex VARCHAR, people_id VARCHAR)\n\n question: which gender got the highest average uncertain ratio. ",
55
" Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_name_60 (pick INTEGER, former_wnba_team VARCHAR)\n\n question: What pick was a player that previously played for the Minnesota Lynx? ",
56
- "\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the womens doubles for werner schlager [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the womens doubles for werner schlager [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the womens doubles for werner schlager [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE",
56
+ "\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the women's doubles for werner schlager [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the women's doubles for werner schlager [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE table_28138035_4 (womens_doubles VARCHAR, mens_singles VARCHAR)\n\n question: Name the women's doubles for werner schlager [/user] [assistant]\n\n [user] Write a SQL query to answer the question based on the table schema.\n\n context: CREATE TABLE",
57
]
58
expected_lora_output = [
59
" SELECT icao FROM table_name_74 WHERE airport = 'lilongwe international airport' ",
vllm/core/block_manager.py
+1
-1
@@ -178,7 +178,7 @@ class BlockSpaceManager:
178
if len(block_table) < len(logical_blocks):
179
if (self.block_sliding_window
180
and len(block_table) >= self.block_sliding_window):
181
- # re-use a block
181
+ # reuse a block
182
block_table.append(block_table[len(block_table) %
183
self.block_sliding_window])
184
else:
vllm/core/scheduler.py
+1
-1
@@ -158,7 +158,7 @@ class Scheduler:
158
return len(self.waiting) + len(self.running) + len(self.swapped)
159
160
def _schedule(self) -> SchedulerOutputs:
161
- # Blocks that need to be swaped or copied before model execution.
161
+ # Blocks that need to be swapped or copied before model execution.
162
blocks_to_swap_in: Dict[int, int] = {}
163
blocks_to_swap_out: Dict[int, int] = {}
164
blocks_to_copy: Dict[int, List[int]] = {}
vllm/lora/punica.py
+1
-1
@@ -87,7 +87,7 @@ def add_lora(y: torch.Tensor,
87
r = wb_t_all.size(-1)
88
if buffer is None:
89
# We set the buffer to be float32 by default to avoid
90
- # numerical innacuracies that would otherwise happen
90
+ # numerical inaccuracies that would otherwise happen
91
# due to downcasting.
92
buffer = torch.zeros((x.size(0), r),
93
dtype=torch.float32,
vllm/model_executor/layers/triton_kernel/prefix_prefill.py
+1
-1
@@ -537,7 +537,7 @@ if triton.__version__ >= "2.1.0":
537
alibi_start_q = tl.arange(
538
0, BLOCK_M) + block_start_loc + cur_batch_ctx_len
539
alibi_start_k = cur_batch_ctx_len
540
- # # init debuger
540
+ # # init debugger
541
# offset_db_q = tl.arange(0, BLOCK_M) + block_start_loc
542
# offset_db_k = tl.arange(0, BLOCK_N)
543
# calc q[BLOCK_M, BLOCK_MODEL] mul k[prefix_len: , BLOCK_DMODEL]
vllm/model_executor/models/decilm.py
+1
-1
@@ -41,7 +41,7 @@ class DeciLMForCausalLM(LlamaForCausalLM):
41
Based on the llama executor.
42
43
The main difference is that DeciLM uses Variable Grouped Query Attention.
44
- The constant number of GQA heads in the decoder is overriden with a value
44
+ The constant number of GQA heads in the decoder is overridden with a value
45
per layer.
46
47
Usually, in the HuggingFace implementation, instead of
vllm/model_executor/parallel_utils/custom_all_reduce.py
+2
-2
@@ -36,14 +36,14 @@ def init_custom_ar() -> None:
36
if world_size not in _SUPPORTED_WORLD_SIZES:
37
logger.warn(
38
"Custom allreduce is disabled due to an unsupported world size: "
39
- "%d. Supported world sizes: %s. To slience this warning, specify"
39
+ "%d. Supported world sizes: %s. To silence this warning, specify"
40
"disable_custom_all_reduce=True explicitly.", world_size,
41
str(_SUPPORTED_WORLD_SIZES))
42
return
43
if not _can_p2p(rank, world_size):
44
logger.warn(
45
"Custom allreduce is disabled because your platform lacks GPU P2P"
46
- " capability. To slience this warning, specify"
46
+ " capability. To silence this warning, specify"
47
"disable_custom_all_reduce=True explicitly.")
48
return
49
_CA_HANDLE = CustomAllreduce(rank, world_size)
vllm/model_executor/parallel_utils/parallel_state.py
+1
-1
@@ -189,7 +189,7 @@ def get_pipeline_model_parallel_next_rank():
189
190
191
def get_pipeline_model_parallel_prev_rank():
192
- """Return the global rank that preceeds the caller in the pipeline"""
192
+ """Return the global rank that precedes the caller in the pipeline"""
193
assert _PIPELINE_GLOBAL_RANKS is not None, (
194
"Pipeline parallel group is not initialized")
195
rank_in_pipeline = get_pipeline_model_parallel_rank()
vllm/utils.py
+1
-1
@@ -204,7 +204,7 @@ def _generate_random_fp8_e5m2(
204
# NOTE(zhaoyang): Due to NaN and Inf representation for fp8 data type,
205
# it may occur Inf or NaN if we directly use torch.randint
206
# to generate random data for fp8 data.
207
- # For example, s.11111.00 in fp8e5m2 format repesents Inf.
207
+ # For example, s.11111.00 in fp8e5m2 format represents Inf.
208
# | E4M3 | E5M2
209
#-----|-------------|-------------------
210
# Inf | N/A | s.11111.00