[MyPy][3/N] Fix MyPy errors in test groups (part 3) (#55939)
Signed-off-by: Taimys <ffy2281693949@gmail.com> Signed-off-by: Luyi Xiao <81631911+luyixiao95@users.noreply.github.com> Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Codex <noreply@openai.com> Co-authored-by: Luyi Xiao <81631911+luyixiao95@users.noreply.github.com> Co-authored-by: Taimys <ffy2281693949@gmail.com> Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Codex
Luyi Xiao
Taimys
Harry Mellor
Claude Opus 5.5
parent
c4fa6f36a7
commit
5225f1bacb
@@ -341,6 +341,7 @@ def test_batch_allowed_needs_one_shared_set():
|
||||
assert states.batch_allowed([]) is None
|
||||
|
||||
shared = states.batch_allowed([0])
|
||||
assert shared is not None
|
||||
assert shared.tolist() == [1, 2, 3]
|
||||
assert shared.dtype == torch.int64
|
||||
assert states.batch_allowed([0, 0]) is shared
|
||||
|
||||
@@ -34,6 +34,8 @@ def test_stats_match_reference(vocab: int, dtype: torch.dtype):
|
||||
)
|
||||
assert torch.equal(argmax, ref_argmax)
|
||||
torch.testing.assert_close(entropy, ref_entropy, atol=2e-3, rtol=1e-3)
|
||||
assert probs is not None
|
||||
assert ref_probs is not None
|
||||
torch.testing.assert_close(probs.float(), ref_probs.float(), atol=5e-3, rtol=2e-2)
|
||||
assert probs.dtype == torch.bfloat16
|
||||
assert probs.shape == logits.shape
|
||||
@@ -100,6 +102,7 @@ def test_masked_logits_keep_a_finite_entropy():
|
||||
assert entropy[0].item() < 1e-6 and entropy[1].item() < 1e-6
|
||||
assert abs(entropy[2].item() - math.log(2.0)) < 1e-4
|
||||
assert torch.equal(argmax[[0, 1, 3]], torch.tensor([3, 3, 3], device="cuda"))
|
||||
assert probs is not None
|
||||
assert probs[0, 3].item() == 1.0 and probs[0].sum().item() == 1.0
|
||||
_, _, ref_entropy, _ = sample_row_stats_reference(logits, temps, 1, None)
|
||||
assert torch.isfinite(ref_entropy).all()
|
||||
|
||||
@@ -340,7 +340,7 @@ def _make_v2_state(
|
||||
device=device,
|
||||
all_token_ids=SimpleNamespace(gpu=all_tokens),
|
||||
)
|
||||
state = DryState(_make_vllm_config(), req_states)
|
||||
state = DryState(_make_vllm_config(), req_states) # type: ignore[arg-type]
|
||||
if tokenizer is not None:
|
||||
# Re-resolve the default set now that a tokenizer is installed.
|
||||
state._tokenizer = tokenizer
|
||||
@@ -446,7 +446,7 @@ def test_spec_decode_is_refused_at_construction():
|
||||
)
|
||||
config = _make_vllm_config(speculative_config=SimpleNamespace())
|
||||
with pytest.raises(ValueError, match="speculative decoding"):
|
||||
DryState(config, req_states)
|
||||
DryState(config, req_states) # type: ignore[arg-type]
|
||||
|
||||
|
||||
def test_v2_draft_expanded_logits_raise():
|
||||
@@ -917,7 +917,10 @@ def test_loads_through_the_loader_by_fqcn():
|
||||
total_len=None,
|
||||
)
|
||||
procs = loader.build_custom_logits_processors(
|
||||
_make_vllm_config(), req_states, False, [DRY_FQCN]
|
||||
_make_vllm_config(),
|
||||
req_states, # type: ignore[arg-type]
|
||||
False,
|
||||
[DRY_FQCN],
|
||||
)
|
||||
assert [type(p) for p in procs] == [DryState]
|
||||
|
||||
|
||||
@@ -16,6 +16,7 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
||||
from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
ParallelLMHead,
|
||||
UnquantizedEmbeddingMethod,
|
||||
VocabParallelEmbedding,
|
||||
)
|
||||
|
||||
|
||||
@@ -32,10 +33,17 @@ class _FakeLmHead:
|
||||
self.tp_size = 1
|
||||
|
||||
|
||||
def _fake_lm_head(
|
||||
weight: torch.Tensor, quantized: bool = False, shard_indices: object | None = None
|
||||
) -> VocabParallelEmbedding:
|
||||
# The dtype tests only need the projection fields, not distributed setup.
|
||||
return _FakeLmHead(weight, quantized, shard_indices) # type: ignore[return-value]
|
||||
|
||||
|
||||
def _build_processor(vocab_size: int) -> LogitsProcessor:
|
||||
lp = LogitsProcessor(vocab_size)
|
||||
# The TP gather is orthogonal to the dtype behavior under test.
|
||||
lp._gather_logits = lambda logits: logits
|
||||
lp._gather_logits = lambda logits: logits # type: ignore[method-assign]
|
||||
return lp
|
||||
|
||||
|
||||
@@ -47,8 +55,9 @@ def test_fp32_head_runs_projection_in_fp32(default_vllm_config):
|
||||
hidden_states = torch.randn(num_tokens, hidden_size, dtype=torch.bfloat16)
|
||||
weight = torch.randn(vocab_size, hidden_size, dtype=torch.bfloat16)
|
||||
|
||||
logits = lp._get_logits(hidden_states, _FakeLmHead(weight), None)
|
||||
logits = lp._get_logits(hidden_states, _fake_lm_head(weight), None)
|
||||
|
||||
assert logits is not None
|
||||
assert logits.dtype == torch.float32
|
||||
assert torch.isfinite(logits).all()
|
||||
expected = torch.nn.functional.linear(hidden_states.float(), weight.float())
|
||||
@@ -66,8 +75,9 @@ def test_non_fp32_head_dtype_uses_cast_path(default_vllm_config):
|
||||
hidden_states = torch.randn(4, hidden_size, dtype=torch.bfloat16)
|
||||
weight = torch.randn(vocab_size, hidden_size, dtype=torch.bfloat16)
|
||||
|
||||
logits = lp._get_logits(hidden_states, _FakeLmHead(weight), None)
|
||||
logits = lp._get_logits(hidden_states, _fake_lm_head(weight), None)
|
||||
|
||||
assert logits is not None
|
||||
assert logits.dtype == torch.float16
|
||||
expected = torch.nn.functional.linear(hidden_states.half(), weight.half())
|
||||
torch.testing.assert_close(logits, expected)
|
||||
@@ -82,7 +92,7 @@ def test_head_dtype_equal_to_model_dtype_uses_quant_method(default_vllm_config):
|
||||
|
||||
hidden_states = torch.randn(4, hidden_size, dtype=torch.bfloat16)
|
||||
weight = torch.randn(vocab_size, hidden_size, dtype=torch.bfloat16)
|
||||
lm_head = _FakeLmHead(weight)
|
||||
lm_head = _fake_lm_head(weight)
|
||||
|
||||
with mock.patch.object(
|
||||
lm_head.quant_method,
|
||||
@@ -94,6 +104,7 @@ def test_head_dtype_equal_to_model_dtype_uses_quant_method(default_vllm_config):
|
||||
logits = lp._get_logits(hidden_states, lm_head, None)
|
||||
|
||||
apply_mock.assert_called_once()
|
||||
assert logits is not None
|
||||
assert logits.dtype == torch.bfloat16
|
||||
|
||||
|
||||
@@ -120,9 +131,10 @@ def test_fp32_head_uses_mm_fast_path_on_device(default_vllm_config):
|
||||
with mock.patch(
|
||||
"vllm.model_executor.layers.logits_processor.F.linear"
|
||||
) as linear_mock:
|
||||
logits = lp._get_logits(hidden_states, _FakeLmHead(weight), None)
|
||||
logits = lp._get_logits(hidden_states, _fake_lm_head(weight), None)
|
||||
|
||||
linear_mock.assert_not_called()
|
||||
assert logits is not None
|
||||
assert logits.dtype == torch.float32
|
||||
expected = torch.nn.functional.linear(hidden_states.float(), weight.float())
|
||||
torch.testing.assert_close(logits, expected)
|
||||
@@ -131,7 +143,7 @@ def test_fp32_head_uses_mm_fast_path_on_device(default_vllm_config):
|
||||
def test_fp32_head_rejects_quantized_lm_head(default_vllm_config):
|
||||
lp = _build_processor(64)
|
||||
lp.head_dtype = torch.float32
|
||||
lm_head = _FakeLmHead(torch.randn(64, 16, dtype=torch.bfloat16), quantized=True)
|
||||
lm_head = _fake_lm_head(torch.randn(64, 16, dtype=torch.bfloat16), quantized=True)
|
||||
|
||||
with pytest.raises(ValueError, match="unquantized"):
|
||||
lp._get_logits(torch.randn(4, 16, dtype=torch.bfloat16), lm_head, None)
|
||||
@@ -197,7 +209,7 @@ def test_get_top_tokens_honors_head_dtype(default_vllm_config):
|
||||
|
||||
hidden_states = torch.randn(4, hidden_size, dtype=torch.bfloat16)
|
||||
weight = torch.randn(vocab_size, hidden_size, dtype=torch.bfloat16)
|
||||
lm_head = _FakeLmHead(
|
||||
lm_head = _fake_lm_head(
|
||||
weight,
|
||||
shard_indices=types.SimpleNamespace(
|
||||
num_org_vocab_padding=0, org_vocab_start_index=0
|
||||
@@ -236,7 +248,9 @@ def test_fp32_head_e2e_no_nan():
|
||||
|
||||
for output in outputs:
|
||||
for completion in output.outputs:
|
||||
assert completion.logprobs is not None
|
||||
for token_id, position in zip(completion.token_ids, completion.logprobs):
|
||||
assert position is not None
|
||||
# The sampled token survived filtering, so its logprob is finite.
|
||||
assert math.isfinite(position[token_id].logprob)
|
||||
# No returned logprob is NaN.
|
||||
|
||||
@@ -24,6 +24,7 @@ from vllm import SamplingParams
|
||||
from vllm.config.model import LogprobsMode
|
||||
from vllm.distributed import cleanup_dist_env_and_memory
|
||||
from vllm.exceptions import VLLMValidationError
|
||||
from vllm.logprobs import Logprob
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.v1.engine.input_processor import InputProcessor
|
||||
|
||||
@@ -145,7 +146,7 @@ def _repeat_logprob_config(
|
||||
def _run_and_validate(
|
||||
vllm_model: VllmRunner,
|
||||
test_prompts: list[str],
|
||||
vllm_sampling_params: SamplingParams,
|
||||
vllm_sampling_params: SamplingParams | list[SamplingParams],
|
||||
hf_logprobs: list[list[torch.Tensor]],
|
||||
hf_outputs: list[tuple[list[int], str]],
|
||||
logprob_prompt_logprob_list: BatchLogprobsSpecType,
|
||||
@@ -163,11 +164,12 @@ def _run_and_validate(
|
||||
# Extract request-level (prompt)logprobs config
|
||||
num_top_logprobs, num_top_prompt_logprobs = logprob_prompt_logprob
|
||||
|
||||
assert vllm_result.prompt_token_ids is not None
|
||||
# Test whether sampled token output is consistent between vLLM and HF
|
||||
# vLLM prompt+completion should match HF output
|
||||
if temperature == 0.0:
|
||||
assert (
|
||||
vllm_result.prompt_token_ids + vllm_result.outputs[0].token_ids
|
||||
vllm_result.prompt_token_ids + list(vllm_result.outputs[0].token_ids)
|
||||
== hf_output[0]
|
||||
)
|
||||
else:
|
||||
@@ -191,7 +193,9 @@ def _run_and_validate(
|
||||
|
||||
# Confirm that the output token appears among the logprobs
|
||||
assert token_id in logprobs
|
||||
token_in_topk = logprobs[token_id].rank <= num_top_logprobs
|
||||
rank = logprobs[token_id].rank
|
||||
assert rank is not None
|
||||
token_in_topk = rank <= num_top_logprobs
|
||||
|
||||
# If the output token is not included in the top K
|
||||
# logprob, it can return 1 more data
|
||||
@@ -207,8 +211,10 @@ def _run_and_validate(
|
||||
|
||||
output_text = vllm_result.outputs[0].text
|
||||
output_string_from_most_likely_tokens_lst: list[str] = []
|
||||
for top_logprobs in vllm_result.outputs[0].logprobs:
|
||||
top_logprob = next(iter(top_logprobs.values()))
|
||||
for output_logprobs in vllm_result.outputs[0].logprobs:
|
||||
assert output_logprobs is not None
|
||||
top_logprob = next(iter(output_logprobs.values()))
|
||||
assert top_logprob.decoded_token is not None
|
||||
output_string_from_most_likely_tokens_lst.append(
|
||||
top_logprob.decoded_token
|
||||
)
|
||||
@@ -226,8 +232,9 @@ def _run_and_validate(
|
||||
|
||||
# Compare vLLM sample logprobs to HF
|
||||
vllm_sample_logprobs = vllm_result.outputs[0].logprobs
|
||||
for i, top_logprobs in enumerate(vllm_sample_logprobs):
|
||||
for token_id, sample_logprob in top_logprobs.items():
|
||||
for i, sample_logprobs in enumerate(vllm_sample_logprobs):
|
||||
assert sample_logprobs is not None
|
||||
for token_id, sample_logprob in sample_logprobs.items():
|
||||
if temperature == 0.0 or i == 0:
|
||||
logprob = sample_logprob.logprob
|
||||
torch.testing.assert_close(
|
||||
@@ -271,9 +278,9 @@ def _run_and_validate(
|
||||
|
||||
# Confirm that the prompt token appears among the logprobs
|
||||
assert prompt_token_id in prompt_logprobs
|
||||
token_in_topk = (
|
||||
prompt_logprobs[prompt_token_id].rank <= num_top_prompt_logprobs
|
||||
)
|
||||
prompt_rank = prompt_logprobs[prompt_token_id].rank
|
||||
assert prompt_rank is not None
|
||||
token_in_topk = prompt_rank <= num_top_prompt_logprobs
|
||||
|
||||
# If the prompt token is not included in the top K
|
||||
# logprob, it can return 1 more data
|
||||
@@ -294,9 +301,10 @@ def _run_and_validate(
|
||||
# 1:.
|
||||
vllm_prompt_logprobs = vllm_result.prompt_logprobs[1:]
|
||||
for i, vllm_prompt_logprob_dict in enumerate(vllm_prompt_logprobs):
|
||||
for token_id, logprob in vllm_prompt_logprob_dict.items():
|
||||
assert vllm_prompt_logprob_dict is not None
|
||||
for token_id, prompt_lp in vllm_prompt_logprob_dict.items():
|
||||
torch.testing.assert_close(
|
||||
logprob.logprob,
|
||||
prompt_lp.logprob,
|
||||
hf_logprob[0][i][token_id].item(),
|
||||
atol=2e-2,
|
||||
rtol=2e-2,
|
||||
@@ -478,12 +486,12 @@ def test_prompt_logprob_token_ids_require_v2_model_runner():
|
||||
params = SamplingParams(prompt_logprob_token_ids=[1, 2])
|
||||
with patch.object(SamplingParams, "verify"):
|
||||
with pytest.raises(VLLMValidationError, match="V2 model runner"):
|
||||
InputProcessor._validate_params(processor, params, ("generate",))
|
||||
InputProcessor._validate_params(processor, params, ("generate",)) # type: ignore[arg-type]
|
||||
processor.vllm_config.use_v2_model_runner = True
|
||||
InputProcessor._validate_params(processor, params, ("generate",))
|
||||
InputProcessor._validate_params(processor, params, ("generate",)) # type: ignore[arg-type]
|
||||
processor.vllm_config.cache_config.kv_sharing_fast_prefill = True
|
||||
with pytest.raises(VLLMValidationError, match="fast-prefill"):
|
||||
InputProcessor._validate_params(processor, params, ("generate",))
|
||||
InputProcessor._validate_params(processor, params, ("generate",)) # type: ignore[arg-type]
|
||||
|
||||
|
||||
def test_none_logprobs(vllm_model, example_prompts):
|
||||
@@ -607,7 +615,9 @@ def test_logprobs_mode(logprobs_mode: LogprobsMode):
|
||||
total_token_with_logprobs = 0
|
||||
positive_values = 0
|
||||
for output in results[0].outputs:
|
||||
assert output.logprobs is not None
|
||||
for logprobs in output.logprobs:
|
||||
assert logprobs is not None
|
||||
for token_id in logprobs:
|
||||
logprob = logprobs[token_id]
|
||||
if logprobs_mode in ("raw_logprobs", "processed_logprobs"):
|
||||
@@ -648,6 +658,7 @@ def test_prompt_logprobs_mode():
|
||||
)
|
||||
assert results[0].prompt_logprobs is not None
|
||||
assert results[0].prompt_logprobs[1] is not None
|
||||
assert results[0].prompt_token_ids is not None
|
||||
tok_id = results[0].prompt_token_ids[1]
|
||||
values[mode] = results[0].prompt_logprobs[1][tok_id].logprob
|
||||
finally:
|
||||
@@ -1228,10 +1239,12 @@ def test_spec_decode_logprobs(
|
||||
[prompt, prompt], [sampling_params, penalty_sampling_params]
|
||||
)
|
||||
# Collect logprobs outputs from reference LLM.
|
||||
ref_logprobs = []
|
||||
ref_logprobs: list[Logprob] = []
|
||||
for results in ref_results:
|
||||
for output in results.outputs:
|
||||
assert output.logprobs is not None
|
||||
for logprobs in output.logprobs:
|
||||
assert logprobs is not None
|
||||
ref_logprobs.extend(logprobs.values())
|
||||
del ref_llm
|
||||
torch.accelerator.empty_cache()
|
||||
@@ -1249,10 +1262,12 @@ def test_spec_decode_logprobs(
|
||||
[prompt, prompt], [sampling_params, penalty_sampling_params]
|
||||
)
|
||||
# Collect logprobs outputs from spec decode LLM.
|
||||
spec_logprobs = []
|
||||
spec_logprobs: list[Logprob] = []
|
||||
for results in spec_results:
|
||||
for output in results.outputs:
|
||||
assert output.logprobs is not None
|
||||
for logprobs in output.logprobs:
|
||||
assert logprobs is not None
|
||||
spec_logprobs.extend(logprobs.values())
|
||||
del spec_llm
|
||||
torch.accelerator.empty_cache()
|
||||
@@ -1326,6 +1341,7 @@ def test_prompt_logprobs_with_chunking_and_preemption():
|
||||
"Unexpected number of prompt logprob positions"
|
||||
)
|
||||
|
||||
assert sampling_params.prompt_logprobs is not None
|
||||
# Each position should have the requested number of logprobs
|
||||
for pos, logprobs_dict in enumerate(prompt_logprobs):
|
||||
if logprobs_dict is not None: # First token may be None
|
||||
|
||||
@@ -333,6 +333,7 @@ def test_sampler_repetition_penalty(
|
||||
for batch_idx in range(batch_size):
|
||||
non_penalized_token_id = logits[batch_idx].argmax().item()
|
||||
penalized_token_id = logits[batch_idx].argmin().item()
|
||||
assert sampling_metadata.prompt_token_ids is not None
|
||||
prompt_tokens = sampling_metadata.prompt_token_ids[batch_idx][:].tolist()
|
||||
output_tokens = sampling_metadata.output_token_ids[batch_idx]
|
||||
if repetition_penalty > 1.0:
|
||||
|
||||
@@ -4,6 +4,7 @@
|
||||
|
||||
import torch
|
||||
|
||||
from tests.v1.sample.utils import create_mock_reasoning_config
|
||||
from vllm.sampling_params import SamplingParams
|
||||
from vllm.v1.sample.logits_processor.interface import (
|
||||
BatchUpdate,
|
||||
@@ -12,14 +13,9 @@ from vllm.v1.sample.logits_processor.interface import (
|
||||
from vllm.v1.sample.thinking_budget_state import ThinkingBudgetStateHolder
|
||||
|
||||
|
||||
class _MockReasoningConfig:
|
||||
reasoning_start_token_ids = [151667]
|
||||
reasoning_end_token_ids = [151668]
|
||||
|
||||
|
||||
def _make_holder() -> ThinkingBudgetStateHolder:
|
||||
return ThinkingBudgetStateHolder(
|
||||
_MockReasoningConfig(),
|
||||
create_mock_reasoning_config([151667], [151668]),
|
||||
8,
|
||||
0,
|
||||
torch.device("cpu"),
|
||||
|
||||
@@ -104,7 +104,8 @@ def test_trace_decode_token_ids_rejects_speculative_decoding():
|
||||
params = SamplingParams(trace_decode_token_ids=[1])
|
||||
with pytest.raises(ValueError, match="not supported with speculative decoding"):
|
||||
params._validate_trace_replay(
|
||||
_make_model_config(vocab_size=100), speculative_config=object()
|
||||
_make_model_config(vocab_size=100),
|
||||
speculative_config=object(), # type: ignore[arg-type] # Only non-None presence is tested.
|
||||
)
|
||||
|
||||
|
||||
|
||||
@@ -4,16 +4,28 @@
|
||||
from collections.abc import Iterator
|
||||
from enum import Enum
|
||||
from typing import NamedTuple
|
||||
from unittest.mock import MagicMock
|
||||
|
||||
import regex as re
|
||||
import torch
|
||||
|
||||
from vllm import CompletionOutput
|
||||
from vllm.config import ReasoningConfig
|
||||
from vllm.utils.torch_utils import make_tensor_with_pad
|
||||
from vllm.v1.sample.logits_processor import BatchUpdate, LogitsProcessor
|
||||
from vllm.v1.sample.metadata import SamplingMetadata
|
||||
|
||||
|
||||
def create_mock_reasoning_config(
|
||||
start_token_ids: list[int], end_token_ids: list[int]
|
||||
) -> ReasoningConfig:
|
||||
return MagicMock(
|
||||
spec=ReasoningConfig,
|
||||
reasoning_start_token_ids=start_token_ids,
|
||||
reasoning_end_token_ids=end_token_ids,
|
||||
)
|
||||
|
||||
|
||||
class BatchLogprobsComposition(Enum):
|
||||
"""Types of logprobs configs to include in test batch."""
|
||||
|
||||
@@ -141,7 +153,11 @@ def compute_correct_cumulative_logprob(completion_output: CompletionOutput) -> f
|
||||
token_ids = completion_output.token_ids
|
||||
logprobs = completion_output.logprobs
|
||||
assert logprobs is not None
|
||||
return sum([lp[tok_id].logprob for tok_id, lp in zip(token_ids, logprobs)])
|
||||
values: list[float] = []
|
||||
for tok_id, lp in zip(token_ids, logprobs):
|
||||
assert lp is not None
|
||||
values.append(lp[tok_id].logprob)
|
||||
return sum(values)
|
||||
|
||||
|
||||
def create_fake_logits(batch_size: int, vocab_size: int) -> torch.Tensor:
|
||||
|
||||
Reference in New Issue
Block a user