environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_USE_HPU_CONTIGUOUS_CACHE_FETCH": lambda: (
os.environ.get("VLLM_CONTIGUOUS_PA", "true").lower()
in ("1", "true")
),
"VLLM_HPU_FORCE_CHANNEL_FP8": lambda: (
os.environ.get(
"VLLM_HPU_FORCE_CHANNEL_FP8", "true"
).lower()
in ("1", "true")
and os.environ.get("QUANT_CONFIG", None) is None
),
"VLLM_HPU_HETERO_KV_LAYOUT": lambda: (
os.environ.get(
"VLLM_HPU_HETERO_KV_LAYOUT", "false"
).lower()
in ("0", "false")
),
"VLLM_HPU_MULTI_MODEL_CONFIG": lambda: os.environ.get(
"VLLM_HPU_MULTI_MODEL_CONFIG", None
),
"VLLM_NIXL_ABORT_REQUEST_TIMEOUT": lambda: float(
os.environ.get(
"VLLM_NIXL_ABORT_REQUEST_TIMEOUT", "300"
)
),
"VLLM_NIXL_SIDE_CHANNEL_HOST": lambda: os.environ.get(
"VLLM_NIXL_SIDE_CHANNEL_HOST", "localhost"
),
"VLLM_NIXL_SIDE_CHANNEL_PORT": lambda: int(
os.environ.get(
"VLLM_NIXL_SIDE_CHANNEL_PORT", "5600"
)
),
"VLLM_HPU_NIXL_JOINT_KV": lambda: (
os.environ.get(
"VLLM_HPU_NIXL_JOINT_KV", "false"
).lower()
in ("1", "true")
),
"VLLM_HPU_NIXL_STAGING_SLOTS": lambda: int(
os.environ.get("VLLM_HPU_NIXL_STAGING_SLOTS", "0")
),
}