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

[Cleanup] Remove obsolete spec decoding compatibility logic (#32003)

Browse Source 
Signed-off-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Nick Hill
2026-01-08 21:44:18 -08:00
committed by GitHub
parent 7a05d2dc65
commit 29ce48221c
8 changed files with 45 additions and 75 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
tests/v1/sample/test_logprobs.py
View File

@@ -940,27 +940,62 @@ def test_correct_decoded_token_preserves_valid_tokens():
(
"eagle",
"meta-llama/Llama-3.2-1B-Instruct",
"nm-testing/Llama3_2_1B_speculator.eagle3",
{
"method": "eagle",
"model": "nm-testing/Llama3_2_1B_speculator.eagle3",
"num_speculative_tokens": 3,
},
0,
),
marks=large_gpu_mark(min_gb=32),
id="eagle0",
),
pytest.param(
(
"eagle",
"meta-llama/Llama-3.2-1B-Instruct",
{
"method": "eagle",
"model": "nm-testing/Llama3_2_1B_speculator.eagle3",
"num_speculative_tokens": 3,
},
3,
),
marks=large_gpu_mark(min_gb=32),
id="eagle3",
),
pytest.param(
(
"ngram",
"meta-llama/Llama-3.2-1B-Instruct",
{
"method": "ngram",
"prompt_lookup_max": 5,
"prompt_lookup_min": 3,
"num_speculative_tokens": 3,
},
3,
),
marks=large_gpu_mark(min_gb=32),
id="ngram",
),
],
)
@pytest.mark.parametrize("top_logprobs", [0, 3])
def test_spec_decode_logprobs(
logprobs_mode: LogprobsMode,
model_setup: tuple[str, str, str],
top_logprobs: int,
model_setup: tuple[str, str, dict, int],
):
"""Spec decode logprobs should match those of the base model.
Args:
logprobs_mode: logprobs mode.
model_setup: Spec decode method, base model name, and
draft model name.
model_setup: Tuple of (method, base model name,
speculative_config dict, top_logprobs).
"""
from vllm import LLM
method, model_name, spec_config, top_logprobs = model_setup
prompt = "Hello world " * 50
sampling_params = SamplingParams(
temperature=0, logprobs=top_logprobs, max_tokens=10, ignore_eos=False
@@ -972,7 +1007,7 @@ def test_spec_decode_logprobs(
ignore_eos=False,
presence_penalty=-1.0,
)
method, model_name, spec_model_name = model_setup
max_model_len = 256
# Run base LLM.
@@ -999,14 +1034,11 @@ def test_spec_decode_logprobs(
cleanup_dist_env_and_memory()
# Run spec decode LLM.
# Add max_model_len to spec_config if not present
spec_config_with_len = {**spec_config, "max_model_len": max_model_len}
spec_llm = LLM(
model_name,
speculative_config={
"method": method,
"model": spec_model_name,
"num_speculative_tokens": 3,
"max_model_len": max_model_len,
},
speculative_config=spec_config_with_len,
max_logprobs=5,
max_model_len=max_model_len,
seed=42,

20
tests/v1/spec_decode/test_ngram.py
View File

@@ -82,10 +82,8 @@ def test_ngram_proposer():
token_ids_cpu = np.array([[1, 2, 3, 4, 5]])
result = get_ngram_proposer(min_n=2, max_n=2, k=2).propose(
sampled_token_ids=[[0]],
req_ids=["0"],
num_tokens_no_spec=np.array([len(c) for c in token_ids_cpu]),
token_ids_cpu=token_ids_cpu,
spec_decode_unsupported_reqs=(),
)
assert len(result[0]) == 0
@@ -93,10 +91,8 @@ def test_ngram_proposer():
token_ids_cpu = np.array([[1, 2, 3, 4, 1, 2, 3]])
result = get_ngram_proposer(min_n=4, max_n=4, k=2).propose(
sampled_token_ids=[[0]],
req_ids=["0"],
num_tokens_no_spec=np.array([len(c) for c in token_ids_cpu]),
token_ids_cpu=token_ids_cpu,
spec_decode_unsupported_reqs=(),
)
assert len(result[0]) == 0
@@ -104,10 +100,8 @@ def test_ngram_proposer():
token_ids_cpu = np.array([[1, 2, 3, 4, 1, 2, 3]])
result = get_ngram_proposer(min_n=3, max_n=4, k=2).propose(
sampled_token_ids=[[0]],
req_ids=["0"],
num_tokens_no_spec=np.array([len(c) for c in token_ids_cpu]),
token_ids_cpu=token_ids_cpu,
spec_decode_unsupported_reqs=(),
)
assert np.array_equal(result, np.array([[4, 1]]))
@@ -116,10 +110,8 @@ def test_ngram_proposer():
token_ids_cpu = np.array([[2, 3, 4, 5, 1, 2, 3, 4, 1, 2, 3, 4]])
result = get_ngram_proposer(min_n=3, max_n=4, k=2).propose(
sampled_token_ids=[[0]],
req_ids=["0"],
num_tokens_no_spec=np.array([len(c) for c in token_ids_cpu]),
token_ids_cpu=token_ids_cpu,
spec_decode_unsupported_reqs=(),
)
assert np.array_equal(result, np.array([[1, 2]])) # Not [5, 1]]
@@ -127,10 +119,8 @@ def test_ngram_proposer():
token_ids_cpu = np.array([[3, 4, 5, 2, 3, 4, 1, 2, 3, 4]])
result = get_ngram_proposer(min_n=2, max_n=4, k=2).propose(
sampled_token_ids=[[0]],
req_ids=["0"],
num_tokens_no_spec=np.array([len(c) for c in token_ids_cpu]),
token_ids_cpu=token_ids_cpu,
spec_decode_unsupported_reqs=(),
)
assert np.array_equal(result, np.array([[1, 2]])) # Not [5, 2]]
@@ -138,10 +128,8 @@ def test_ngram_proposer():
token_ids_cpu = np.array([[1, 2, 3, 100, 1, 2, 3, 200, 1, 2, 3, 300, 1, 2, 3]])
result = get_ngram_proposer(min_n=3, max_n=3, k=2).propose(
sampled_token_ids=[[0]],
req_ids=["0"],
num_tokens_no_spec=np.array([len(c) for c in token_ids_cpu]),
token_ids_cpu=token_ids_cpu,
spec_decode_unsupported_reqs=(),
)
assert np.array_equal(result, np.array([[100, 1]]))
@@ -149,10 +137,8 @@ def test_ngram_proposer():
token_ids_cpu = np.array([[]])
result = get_ngram_proposer(min_n=2, max_n=2, k=2).propose(
sampled_token_ids=[[0]],
req_ids=["0"],
num_tokens_no_spec=np.array([len(c) for c in token_ids_cpu]),
token_ids_cpu=token_ids_cpu,
spec_decode_unsupported_reqs=(),
)
assert len(result[0]) == 0
@@ -162,10 +148,8 @@ def test_ngram_proposer():
token_ids_cpu = np.array([[1, 2, 3, 1, 2], [4, 5, 6, -1, -1]])
result = get_ngram_proposer(min_n=2, max_n=2, k=2).propose(
sampled_token_ids=[[0], [1]],
req_ids=["0", "1"],
num_tokens_no_spec=np.array([5, 3]),
token_ids_cpu=token_ids_cpu,
spec_decode_unsupported_reqs=(),
)
assert len(result[0]) == 2
assert np.array_equal(result[0], np.array([3, 1]))
@@ -183,10 +167,8 @@ def test_ngram_proposer():
sampled_token_ids = [[2], [], [8]] # Empty list for request 1 simulates prefill
result = proposer.propose(
sampled_token_ids=sampled_token_ids,
req_ids=["0", "1", "2"],
num_tokens_no_spec=num_tokens_no_spec,
token_ids_cpu=token_ids_cpu,
spec_decode_unsupported_reqs=(),
)
assert len(result) == 3
assert np.array_equal(result[0], [3, 1])
@@ -214,10 +196,8 @@ def test_ngram_proposer():
token_ids_cpu = np.array([input_1, input_2])
result = ngram_proposer.propose(
sampled_token_ids=[[0], [1]],
req_ids=["0", "1"],
num_tokens_no_spec=np.array([len(input_1), 3]),
token_ids_cpu=token_ids_cpu,
spec_decode_unsupported_reqs=(),
)
assert len(result[0]) == 2
assert np.array_equal(result[0], np.array([middle_integer + 2, middle_integer + 3]))