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

[V1] V1 Enablement Oracle (#13726)

Browse Source 
Signed-off-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com>
Co-authored-by: rshaw@neuralmagic.com <rshaw@neuralmagic.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Michael Goin <michael@neuralmagic.com>
This commit is contained in:
Robert Shaw
2025-03-15 01:02:20 -04:00
committed by GitHub
parent 8c0d15d5c5
commit d4d93db2c5
96 changed files with 1537 additions and 512 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
tests/models/decoder_only/language/test_gguf.py
View File

@@ -110,16 +110,6 @@ def test_models(
example_prompts = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True)
# Run unquantized model.
with vllm_runner(
model_name=model.original_model,
enforce_eager=True, # faster tests
dtype=dtype,
max_model_len=MAX_MODEL_LEN,
tensor_parallel_size=tp_size) as original_model:
original_outputs = original_model.generate_greedy_logprobs(
example_prompts[:-1], max_tokens, num_logprobs)
# Run gguf model.
with vllm_runner(model_name=model.gguf_model,
enforce_eager=True,
@@ -130,6 +120,16 @@ def test_models(
gguf_outputs = gguf_model.generate_greedy_logprobs(
example_prompts[:-1], max_tokens, num_logprobs)
# Run unquantized model.
with vllm_runner(
model_name=model.original_model,
enforce_eager=True, # faster tests
dtype=dtype,
max_model_len=MAX_MODEL_LEN,
tensor_parallel_size=tp_size) as original_model:
original_outputs = original_model.generate_greedy_logprobs(
example_prompts[:-1], max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=original_outputs,
outputs_1_lst=gguf_outputs,

14
tests/models/decoder_only/language/test_hybrid.py
View File

@@ -9,7 +9,9 @@ from vllm.sampling_params import SamplingParams
from ...utils import check_outputs_equal
# This test is for the hybrid models
MODELS = ["ai21labs/Jamba-tiny-dev", "ibm-ai-platform/Bamba-9B"]
MODELS = ["ai21labs/Jamba-tiny-dev"]
# Bamba at Fp32 is too big for the CI (L4 GPU).
# MODELS = ["ai21labs/Jamba-tiny-dev", "ibm-ai-platform/Bamba-9B"]
@pytest.mark.parametrize("model", MODELS)
@@ -41,13 +43,6 @@ def test_models(
with vllm_runner(model, dtype=dtype) as vllm_model:
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
# This test is for verifying whether the model's extra_repr
# can be printed correctly.
def print_model(model):
print(model)
vllm_model.apply_model(print_model)
for i in range(len(example_prompts)):
hf_output_ids, hf_output_str = hf_outputs[i]
vllm_output_ids, vllm_output_str = vllm_outputs[i]
@@ -192,6 +187,7 @@ def test_parallel_sampling(
)
@pytest.mark.skip(reason="RE-ENABLE: test is currently failing on main.")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("max_tokens", [20])
@@ -293,6 +289,7 @@ def test_state_cleanup(
"could be related to finished_requests_ids")
@pytest.mark.skip(reason="RE-ENABLE: test is currently failing on main.")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
def test_multistep(
@@ -308,6 +305,7 @@ def test_multistep(
vllm_model.generate_greedy([example_prompts[0]] * 10, 1)
@pytest.mark.skip(reason="RE-ENABLE: test is currently failing on main.")
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["float"])
@pytest.mark.parametrize("max_tokens", [64])

7
tests/models/decoder_only/language/test_mamba.py
View File

@@ -68,13 +68,6 @@ def test_models(
with vllm_runner(model, dtype=dtype, max_num_seqs=16) as vllm_model:
vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens)
# This test is for verifying whether the model's extra_repr
# can be printed correctly.
def print_model(model):
print(model)
vllm_model.apply_model(print_model)
for i in range(len(example_prompts)):
hf_output_ids, hf_output_str = hf_outputs[i]
vllm_output_ids, vllm_output_str = vllm_outputs[i]

21
tests/models/decoder_only/language/test_mistral.py
View File

@@ -213,16 +213,6 @@ def test_mistral_format(
max_tokens: int,
num_logprobs: int,
) -> None:
with vllm_runner(
model,
dtype=dtype,
tokenizer_mode="auto",
load_format="safetensors",
config_format="hf",
) as hf_format_model:
hf_format_outputs = hf_format_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
with vllm_runner(
model,
dtype=dtype,
@@ -233,6 +223,16 @@ def test_mistral_format(
mistral_format_outputs = mistral_format_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
with vllm_runner(
model,
dtype=dtype,
tokenizer_mode="auto",
load_format="safetensors",
config_format="hf",
) as hf_format_model:
hf_format_outputs = hf_format_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
check_logprobs_close(
outputs_0_lst=hf_format_outputs,
outputs_1_lst=mistral_format_outputs,
@@ -261,6 +261,7 @@ def test_mistral_symbolic_languages(
assert "<EFBFBD>" not in outputs[0].outputs[0].text.strip()
@pytest.mark.skip("RE-ENABLE: test is currently failing on main.")
@pytest.mark.parametrize("dtype", ["bfloat16"])
@pytest.mark.parametrize("model",
MISTRAL_FORMAT_MODELS) # v1 can't do func calling

16
tests/models/decoder_only/language/test_models.py
View File

@@ -7,6 +7,12 @@ import pytest
from ...utils import check_logprobs_close
# These have unsupported head_dim for FA. We do not
# not have a clean way to fall back, so we fail with
# a clear msg when it happens.
# https://github.com/vllm-project/vllm/issues/14524
REQUIRES_V0 = ["microsoft/phi-2", "stabilityai/stablelm-3b-4e1t"]
@pytest.mark.parametrize(
"model",
@@ -71,7 +77,10 @@ def test_models(
dtype: str,
max_tokens: int,
num_logprobs: int,
monkeypatch,
) -> None:
if model in REQUIRES_V0:
monkeypatch.setenv("VLLM_USE_V1", "0")
with hf_runner(model, dtype=dtype) as hf_model:
if model.startswith("THUDM/chatglm3"):
@@ -85,13 +94,6 @@ def test_models(
vllm_outputs = vllm_model.generate_greedy_logprobs(
example_prompts, max_tokens, num_logprobs)
# This test is for verifying whether the model's extra_repr
# can be printed correctly.
def print_model(model):
print(model)
vllm_model.apply_model(print_model)
check_logprobs_close(
outputs_0_lst=hf_outputs,
outputs_1_lst=vllm_outputs,