biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CI] change spell checker from codespell to typos (#18711)

Browse Source 
Signed-off-by: Andy Xie <andy.xning@gmail.com>
This commit is contained in:
Ning Xie
2025-06-12 10:57:10 +08:00
committed by GitHub
parent 42f52cc95b
commit 2f1c19b245
57 changed files with 335 additions and 163 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/compile/test_async_tp.py
View File

@@ -223,7 +223,7 @@ def test_async_tp_pass_correctness(
"VLLM_USE_V1": "1",
}
aysnc_tp_args = [
async_tp_args = [
*common_args,
"--tensor-parallel-size",
str(tp_size),
@@ -242,7 +242,7 @@ def test_async_tp_pass_correctness(
]
compare_two_settings(model_id,
aysnc_tp_args,
async_tp_args,
tp_args,
async_tp_env,
tp_env,

4
tests/core/block/e2e/test_correctness.py
View File

@@ -437,8 +437,8 @@ def test_auto_prefix_caching_with_preemption(baseline_llm_generator,
"enable_prefix_caching": True,
}])
@pytest.mark.parametrize("seed", [1])
def test_auto_prefix_caching_after_evition_start(baseline_llm_generator,
test_llm_generator):
def test_auto_prefix_caching_after_eviction_start(baseline_llm_generator,
test_llm_generator):
"""Verify block manager v2 with auto prefix caching could works normal
even when eviction started.
With APC enabled, all blocks are held by native block at the beginning.

6
tests/core/block/e2e/test_correctness_sliding_window.py
View File

@@ -33,8 +33,8 @@ BLOCK_SIZE = 16
@pytest.mark.parametrize("batch_size", [5])
@pytest.mark.parametrize("seed", [1])
@pytest.mark.parametrize("backend", ["FLASH_ATTN", "FLASHINFER", "XFORMERS"])
def test_sliding_window_retrival(baseline_llm_generator, test_llm_generator,
batch_size, seed, backend, monkeypatch):
def test_sliding_window_retrieval(baseline_llm_generator, test_llm_generator,
batch_size, seed, backend, monkeypatch):
"""
The test does a bunch of assignments "x1 = 10\nx2 = 33\n..." and then
asks for value of one of them (which is outside the sliding window).
@@ -100,7 +100,7 @@ def test_sliding_window_retrival(baseline_llm_generator, test_llm_generator,
def test_sliding_window_chunked_prefill(test_llm_generator, batch_size, seed,
backend, monkeypatch):
"""
This is similar to test_sliding_window_retrival, however, it doesn't
This is similar to test_sliding_window_retrieval, however, it doesn't
compare against the v1 block manager since v1 doesn't support
chunked prefill with sliding window.

4
tests/core/test_scheduler.py
View File

@@ -594,8 +594,8 @@ def test_decode_schedule_preempted():
# should be preempted. 1 will also be preempted.
budget = create_token_budget()
output = scheduler._schedule_running(budget, curr_loras)
remainig_running = scheduler.running
assert len(remainig_running) == 0
remaining_running = scheduler.running
assert len(remaining_running) == 0
assert len(output.decode_seq_groups) == 1
assert len(output.prefill_seq_groups) == 0
assert output.decode_seq_groups[0].seq_group.request_id == "0"

4
tests/entrypoints/openai/test_chat_template.py
View File

@@ -16,7 +16,7 @@ chatml_jinja_path = VLLM_PATH / "examples/template_chatml.jinja"
assert chatml_jinja_path.exists()
# Define models, templates, and their corresponding expected outputs
MODEL_TEMPLATE_GENERATON_OUTPUT = [
MODEL_TEMPLATE_GENERATION_OUTPUT = [
("facebook/opt-125m", chatml_jinja_path, True, False, """<|im_start|>user
Hello<|im_end|>
<|im_start|>assistant
@@ -91,7 +91,7 @@ def test_no_load_chat_template_literallike():
@pytest.mark.parametrize(
"model,template,add_generation_prompt,continue_final_message,expected_output",
MODEL_TEMPLATE_GENERATON_OUTPUT)
MODEL_TEMPLATE_GENERATION_OUTPUT)
def test_get_gen_prompt(model, template, add_generation_prompt,
continue_final_message, expected_output):
model_info = HF_EXAMPLE_MODELS.find_hf_info(model)

16
tests/kernels/attention/test_cache.py
View File

@@ -72,8 +72,8 @@ def test_copy_blocks(
# destination blocks.
assert 2 * num_mappings <= num_blocks
src_blocks = random.sample(range(num_blocks), num_mappings)
remainig_blocks = list(set(range(num_blocks)) - set(src_blocks))
dst_blocks = random.sample(remainig_blocks, 2 * num_mappings)
remaining_blocks = list(set(range(num_blocks)) - set(src_blocks))
dst_blocks = random.sample(remaining_blocks, 2 * num_mappings)
block_mapping: list[tuple[int, int]] = []
for i in range(num_mappings):
src = src_blocks[i]
@@ -189,12 +189,12 @@ def test_reshape_and_cache(
# Run the reference implementation.
reshaped_key = key.reshape(num_tokens, *key_cache[0, :, :, 0, :].shape)
block_indicies = torch.div(slot_mapping, block_size, rounding_mode="floor")
block_indicies_lst = block_indicies.cpu().tolist()
block_indices = torch.div(slot_mapping, block_size, rounding_mode="floor")
block_indices_lst = block_indices.cpu().tolist()
block_offsets = slot_mapping % block_size
block_offsets_lst = block_offsets.cpu().tolist()
for i in range(num_tokens):
block_idx = block_indicies_lst[i]
block_idx = block_indices_lst[i]
block_offset = block_offsets_lst[i]
cloned_key_cache[block_idx, :, :, block_offset, :] = reshaped_key[i]
cloned_value_cache[block_idx, :, :, block_offset] = value[i]
@@ -322,12 +322,12 @@ def test_reshape_and_cache_flash(
kv_dtype=kv_cache_dtype)
# Run the reference implementation.
block_indicies = torch.div(slot_mapping, block_size, rounding_mode="floor")
block_indicies_lst = block_indicies.cpu().tolist()
block_indices = torch.div(slot_mapping, block_size, rounding_mode="floor")
block_indices_lst = block_indices.cpu().tolist()
block_offsets = slot_mapping % block_size
block_offsets_lst = block_offsets.cpu().tolist()
for i in range(num_tokens):
block_idx = block_indicies_lst[i]
block_idx = block_indices_lst[i]
block_offset = block_offsets_lst[i]
if kv_cache_layout == "NHD":
cloned_key_cache[block_idx, block_offset, :, :] = key[i]

2
tests/kernels/attention/test_encoder_decoder_attn.py
View File

@@ -46,7 +46,7 @@ CUDA_DEVICE = "cuda:0"
MAX_DEC_SEQ_LENS = [128]
MAX_ENC_SEQ_LENS = [128]
# Narrow teest-cases for unsupported-scenario
# Narrow test-cases for unsupported-scenario
# tests
HEAD_SIZES_FOR_UNSUPP = [HEAD_SIZES[0]]

8
tests/kernels/core/test_rotary_embedding.py
View File

@@ -39,10 +39,10 @@ def rotary_embedding_opcheck(rot,
@pytest.mark.parametrize("head_size", [32, 108])
@pytest.mark.parametrize("seq_len", [11, 1024])
@pytest.mark.parametrize("use_key", [True, False])
@pytest.mark.parametrize("head_stride_is_contingous", [True, False])
@pytest.mark.parametrize("head_stride_is_contiguous", [True, False])
def test_rotary_embedding_opcheck(dist_init, device, max_position,
is_neox_style, rotary_dim, head_size,
seq_len, use_key, head_stride_is_contingous):
seq_len, use_key, head_stride_is_contiguous):
batch_size = 1
base = 10000
num_heads = 7
@@ -52,7 +52,7 @@ def test_rotary_embedding_opcheck(dist_init, device, max_position,
positions = torch.randint(0,
max_position, (batch_size, seq_len),
device=device)
head_stride = head_size + (64 if head_stride_is_contingous else 0)
head_stride = head_size + (64 if head_stride_is_contiguous else 0)
query = torch.randn(batch_size,
seq_len,
@@ -72,7 +72,7 @@ def test_rotary_embedding_opcheck(dist_init, device, max_position,
# if we have a contiguous head stride, test the alternate
# [..., num_heads * head_dim] shape/layout
if head_stride_is_contingous:
if head_stride_is_contiguous:
rotary_embedding_opcheck(
rot, positions, query.flatten(start_dim=-2),
key.flatten(start_dim=-2) if use_key else None)

27
tests/kernels/mamba/test_mamba_ssm_ssd.py
View File

@@ -107,15 +107,15 @@ def generate_random_inputs(batch_size,
return A, dt, X, B, C
def generate_continous_batched_examples(example_lens_by_batch,
num_examples,
full_length,
last_taken,
exhausted,
n_heads,
d_head,
itype,
device='cuda'):
def generate_continuous_batched_examples(example_lens_by_batch,
num_examples,
full_length,
last_taken,
exhausted,
n_heads,
d_head,
itype,
device='cuda'):
# this function generates a random examples of certain length
# and then cut according to "example_lens_by_batch" and feed
@@ -269,11 +269,10 @@ def test_mamba_chunk_scan_cont_batch(d_head, n_heads, seq_len_chunk_size_cases,
exhausted: dict = {} # map: eg -> boolean indicating example is exhausted
states = None
for Y_min, cu_seqlens, seq_idx, (A, dt, X, B,
C) in generate_continous_batched_examples(
cases, num_examples, seqlen,
last_taken, exhausted, n_heads,
d_head, itype):
for Y_min, cu_seqlens, seq_idx, (
A, dt, X, B, C) in generate_continuous_batched_examples(
cases, num_examples, seqlen, last_taken, exhausted, n_heads,
d_head, itype):
chunk_indices, chunk_offsets = \
_query_start_loc_to_chunk_indices_offsets(

0
tests/lora/test_transfomers_model.py → tests/lora/test_transformers_model.py
View File

2
tests/models/language/generation/test_bart.py
View File

@@ -118,7 +118,7 @@ def run_test(
# default to enforce_eager=True if enforce_eager
# is left unspecified. However, the
# VllmRunner test fixture (which wraps around the LLM class) defaults to
# enforce_eager=False (a behavior which a number of already-exisitng
# enforce_eager=False (a behavior which a number of already-existing
# decoder-only unit tests expect), so when testing an encoder/decoder
# model we must explicitly specify enforce_eager=True in the VllmRunner
# constructor.

2
tests/samplers/test_typical_acceptance_sampler.py
View File

@@ -248,7 +248,7 @@ def test_temperature_zero_target_distribution(seed: int, device: str):
size=(batch_size, 1),
dtype=torch.int64)
# The target probaility distribution is a temperature zero distribution
# with zero entroy. Since our draft token ids don't match the probability
# with zero entropy. Since our draft token ids don't match the probability
# 1.0 tokens in the target distribution we will reject all of them and
# fallback to the greedy sampling for selecting 1 token for each sequence.
# Verify the same.

2
tests/spec_decode/e2e/test_eagle_correctness.py
View File

@@ -18,7 +18,7 @@ However, we still need to verify below scenario could be passed:
* Test greedy equality under various number of speculative tokens.
With those tests, we can say at least, EAGLE would not break the
correctess for the target model outputs.
correctness for the target model outputs.
"""
import pytest

2
tests/spec_decode/e2e/test_medusa_correctness.py
View File

@@ -18,7 +18,7 @@ However, we still need to verify below scenario could be passed:
* Test greedy equality under various number of speculative tokens.
With those tests, we can say at least, Medusa would not break the
correctess for the target model outputs.
correctness for the target model outputs.
"""
import pytest

2
tests/spec_decode/e2e/test_mtp_correctness.py
View File

@@ -18,7 +18,7 @@ However, we still need to verify below scenario could be passed:
* Test greedy equality under various number of speculative tokens.
With those tests, we can say at least, mtp would not break the
correctess for the target model outputs.
correctness for the target model outputs.
"""
import pytest

4
tests/spec_decode/e2e/test_ngram_correctness.py
View File

@@ -22,8 +22,8 @@ However, we still need to verify below scenario could be passed:
* Test greedy equality under preemption
* Test greedy equality under various ngram sizes / speculative sizes
With those tests, we can say at least, ngram spec would not break the correctess
for the target model outputs.
With those tests, we can say at least, ngram spec would not break the
correctness for the target model outputs.
"""
import pytest

2
tests/v1/e2e/test_correctness_sliding_window.py
View File

@@ -30,7 +30,7 @@ model_config = {
])
@pytest.mark.parametrize("batch_size", [5])
@pytest.mark.parametrize("seed", [1])
def test_sliding_window_retrival(monkeypatch, model, batch_size, seed):
def test_sliding_window_retrieval(monkeypatch, model, batch_size, seed):
"""
The test does a bunch of assignments "x1 = 10\nx2 = 33\n..." and then
asks for value of one of them (which is outside the sliding window).

4
tests/v1/kv_connector/unit/test_nixl_connector.py
View File

@@ -7,7 +7,7 @@ from vllm.distributed.kv_transfer.kv_connector.v1.nixl_connector import (
from .utils import create_request, create_scheduler, create_vllm_config
def test_basic_inferface():
def test_basic_interface():
"""Unit test for basic NixlConnector interface functionality."""
vllm_config = create_vllm_config()
@@ -25,7 +25,7 @@ def test_basic_inferface():
scheduler.add_request(request)
# Remote Prefill, triggers NixlConnectorMetdata.
# Remote Prefill, triggers NixlConnectorMetadata.
scheduler_output = scheduler.schedule()
kv_connector_metadata = scheduler_output.kv_connector_metadata
assert kv_connector_metadata is not None

2
tests/v1/sample/test_logprobs_e2e.py
View File

@@ -32,7 +32,7 @@ def test_prompt_logprobs_e2e():
), f"Expected: {EXPECTED_VALUE} | Measured: {measured_value}"
def test_promt_logprobs_e2e_server():
def test_prompt_logprobs_e2e_server():
with RemoteOpenAIServer(MODEL, SERVER_ARGS) as remote_server:
url = f"{remote_server.url_for('v1')}/completions"

26
tests/worker/test_model_input.py
View File

@@ -209,32 +209,32 @@ def test_multi_step_model_runner_input():
received_model_input = (StatefulModelInput.from_broadcasted_tensor_dict(
tensor_dict, attn_backend=attn_backend))
receieved_frozen_input = received_model_input.frozen_model_input
received_frozen_input = received_model_input.frozen_model_input
# Check that received copy has correct values.
assert isinstance(received_model_input, StatefulModelInput)
assert receieved_frozen_input.input_tokens is not None
assert (receieved_frozen_input.input_tokens ==
assert received_frozen_input.input_tokens is not None
assert (received_frozen_input.input_tokens ==
frozen_model_input.input_tokens).all()
assert receieved_frozen_input.input_positions is not None
assert (receieved_frozen_input.input_positions ==
assert received_frozen_input.input_positions is not None
assert (received_frozen_input.input_positions ==
frozen_model_input.input_positions).all()
assert receieved_frozen_input.multi_modal_kwargs is None
assert received_frozen_input.multi_modal_kwargs is None
assert (frozen_model_input.multi_modal_kwargs ==
frozen_model_input.multi_modal_kwargs)
assert receieved_frozen_input.lora_requests is None
assert (receieved_frozen_input.lora_requests ==
assert received_frozen_input.lora_requests is None
assert (received_frozen_input.lora_requests ==
frozen_model_input.lora_requests)
assert receieved_frozen_input.lora_mapping is None
assert received_frozen_input.lora_mapping is None
assert (
receieved_frozen_input.lora_mapping == frozen_model_input.lora_mapping)
received_frozen_input.lora_mapping == frozen_model_input.lora_mapping)
for field in dataclasses.fields(AttentionMetadata):
assert getattr(receieved_frozen_input.attn_metadata, field.name,
assert getattr(received_frozen_input.attn_metadata, field.name,
None) == getattr(attn_metadata, field.name, None)
# For sampling metadata, only selected_token_indices is copied.
assert (receieved_frozen_input.sampling_metadata.selected_token_indices ==
assert (received_frozen_input.sampling_metadata.selected_token_indices ==
sampling_metadata.selected_token_indices)
assert receieved_frozen_input.sampling_metadata.seq_groups is None
assert received_frozen_input.sampling_metadata.seq_groups is None
# check non frozen fields
assert received_model_input.is_last_step == model_input.is_last_step