[Engram] Drop redundant VLLM_PLE_CPU_OFFLOAD env var (#57937)
Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
Nicolò Lucchesi committed
Sep 21, 2026 at 20:16 UTC
54020c3c3ec9de219929a5a956010a0251b87520
3 files changed
+5
-23
tests/test_config.py
+2
-10
@@ -1392,16 +1392,8 @@ def test_engram_model_support(
1392
assert resolved.engram_config.cpu_offload is True
1393
1394
1395
-@pytest.mark.parametrize(
1396
- ("value", "expected"), [(None, True), ("0", False), ("1", True)]
1397
-)
1398
-def test_engram_cpu_offload_environment_default(monkeypatch, value, expected):
1399
- monkeypatch.delenv("VLLM_PLE_CPU_OFFLOAD", raising=False)
1400
- if value is not None:
1401
- monkeypatch.setenv("VLLM_PLE_CPU_OFFLOAD", value)
1402
- assert EngramConfig().cpu_offload is expected
1403
- assert EngramConfig(cpu_offload=False).cpu_offload is False
1404
- assert EngramConfig(cpu_offload=True).cpu_offload is True
1395
+def test_engram_cpu_offload_default():
1396
+ assert EngramConfig().cpu_offload is True
1397
1398
1399
def test_engram_config_defaults_to_none():
vllm/config/engram.py
+3
-10
@@ -3,10 +3,9 @@
3
4
from typing import TYPE_CHECKING
5
6
-from pydantic import Field, model_validator
6
+from pydantic import model_validator
7
from typing_extensions import Self
8
9
-import vllm.envs as envs
9
from vllm.config.utils import config, get_hash_factors, hash_factors
10
11
if TYPE_CHECKING:
@@ -22,10 +21,6 @@ _NGRAM_LAYER_FIELDS = {
21
}
22
23
25
-def _default_cpu_offload() -> bool:
26
- return envs.VLLM_PLE_CPU_OFFLOAD
27
-
28
-
24
def model_has_engram_layers(model_config: "ModelConfig | None") -> bool:
25
"""Whether the model carries n-gram embedding layers."""
26
if model_config is None:
@@ -40,10 +35,8 @@ def model_has_engram_layers(model_config: "ModelConfig | None") -> bool:
35
class EngramConfig:
36
"""Configuration for Engram embedding storage and sharding."""
37
43
- cpu_offload: bool = Field(default_factory=_default_cpu_offload)
44
- """Store embedding weights in pinned CPU memory for UVA lookup.
45
- Defaults to VLLM_PLE_CPU_OFFLOAD, which is enabled by default. An explicit
46
- value takes precedence over the environment variable."""
38
+ cpu_offload: bool = True
39
+ """Store embedding weights in pinned CPU memory for UVA lookup."""
40
41
embedding_across_dp: bool = False
42
"""Shard embeddings across TP and all DP ranks when enabled.
vllm/envs.py
-3
@@ -156,7 +156,6 @@ if TYPE_CHECKING:
156
VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT: bool = False
157
VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
158
VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
159
- VLLM_PLE_CPU_OFFLOAD: bool = True
159
VLLM_DISABLE_COMPILE_CACHE: bool = False
160
VLLM_REPLICATE_EMBED: bool = False
161
VLLM_USE_LAYERNAME: bool = True
@@ -2066,8 +2065,6 @@ environment_variables: dict[str, Callable[[], Any]] = {
2065
"VLLM_LOG_MODEL_INSPECTION": lambda: bool(
2066
int(os.getenv("VLLM_LOG_MODEL_INSPECTION", "0"))
2067
),
2069
- # Store n-gram embedding tables in pinned CPU memory for UVA lookup.
2070
- "VLLM_PLE_CPU_OFFLOAD": lambda: bool(int(os.getenv("VLLM_PLE_CPU_OFFLOAD", "1"))),
2068
# Debug logging for --enable-mfu-metrics
2069
"VLLM_DEBUG_MFU_METRICS": lambda: bool(
2070
int(os.getenv("VLLM_DEBUG_MFU_METRICS", "0"))