Fix error in Dynamic NTK scaling (#41277)

Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Signed-off-by: Max de Bayser <maxdebayser@gmail.com>
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
Max de Bayser
2026-05-19 17:27:54 -04:00
committed by GitHub
co-authored by gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> wang.yuqi
parent 1242196295
commit 117afeea46
4 changed files with 45 additions and 152 deletions
@@ -1,7 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# ruff: noqa: SIM117
from typing import Any
import pytest
@@ -40,8 +39,8 @@ def test_default(model_info, vllm_runner):
# For nomic-embed-text-v2-moe the length is set to 512
# by sentence_bert_config.json.
assert model_config.max_model_len == 512
else:
assert model_config.max_model_len == original_max_position_embeddings
if model_info.name == "nomic-ai/nomic-embed-text-v1":
assert model_config.max_model_len == 8192
@pytest.mark.parametrize("model_info", MODELS)
@@ -56,10 +55,9 @@ def test_set_max_model_len_legal(model_info, vllm_runner):
model_config = vllm_model.llm.llm_engine.model_config
assert model_config.max_model_len == 256
# set 512 < max_model_len <= 2048
# For nomic-embed-text-v2-moe the length is set to 512
# by sentence_bert_config.json.
if model_info.name == "nomic-ai/nomic-embed-text-v2-moe":
# For nomic-embed-text-v2-moe the length is set to 512
# by sentence_bert_config.json.
with pytest.raises(ValueError):
with vllm_runner(
model_info.name,
@@ -68,40 +66,27 @@ def test_set_max_model_len_legal(model_info, vllm_runner):
max_model_len=1024,
):
pass
else:
with vllm_runner(
model_info.name,
revision=model_info.revision,
runner="pooling",
max_model_len=1024,
) as vllm_model:
model_config = vllm_model.llm.llm_engine.model_config
assert model_config.max_model_len == 1024
return
# set 512 < max_model_len <= 2048
with vllm_runner(
model_info.name,
revision=model_info.revision,
runner="pooling",
max_model_len=1024,
) as vllm_model:
model_config = vllm_model.llm.llm_engine.model_config
assert model_config.max_model_len == 1024
@pytest.mark.parametrize("model_info", MODELS)
def test_set_max_model_len_illegal(model_info, vllm_runner):
# set max_model_len > 2048
with pytest.raises(ValueError):
with vllm_runner(
model_info.name,
revision=model_info.revision,
runner="pooling",
max_model_len=4096,
):
pass
# set max_model_len > 2048 by hf_overrides
hf_overrides = {"max_model_len": 4096}
with pytest.raises(ValueError):
with vllm_runner(
model_info.name,
revision=model_info.revision,
runner="pooling",
max_model_len=None,
hf_overrides=hf_overrides,
):
pass
with vllm_runner(
model_info.name,
revision=model_info.revision,
runner="pooling",
max_model_len=4096,
) as vllm_model:
model_config = vllm_model.llm.llm_engine.model_config
assert model_config.max_model_len == 4096
@pytest.mark.parametrize("model_info", MODELS)
@@ -124,45 +109,3 @@ def test_use_rope_scaling_legal(model_info, vllm_runner):
hf_overrides=hf_overrides,
):
pass
@pytest.mark.parametrize("model_info", MODELS)
def test_use_rope_scaling_illegal(model_info, vllm_runner):
hf_overrides: dict[str, Any] = {
"rope_parameters": {
"rope_theta": rope_theta,
"rope_type": "yarn",
"factor": factor,
"original_max_position_embeddings": original_max_position_embeddings,
},
}
# illegal max_model_len
with pytest.raises(ValueError):
with vllm_runner(
model_info.name,
revision=model_info.revision,
runner="pooling",
max_model_len=max_model_len + 1,
hf_overrides=hf_overrides,
):
pass
hf_overrides = {
"rope_parameters": {
"rope_theta": rope_theta,
"rope_type": "yarn",
"factor": factor,
"original_max_position_embeddings": original_max_position_embeddings,
},
"max_model_len": max_model_len + 1,
}
# illegal max_model_len by hf_overrides
with pytest.raises(ValueError):
with vllm_runner(
model_info.name,
revision=model_info.revision,
runner="pooling",
max_model_len=None,
hf_overrides=hf_overrides,
):
pass