Replicated embedding and norm fusion for DSV3 flat model (#48484)

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
Jee Jee Li
2026-08-18 14:53:21 +08:00
committed by GitHub
parent 5fa8ca971a
commit 2687fec6ef
5 changed files with 385 additions and 21 deletions
+4
View File
@@ -156,6 +156,7 @@ if TYPE_CHECKING:
VLLM_ENABLE_V1_MULTIPROCESSING: bool = True
VLLM_LOG_BATCHSIZE_INTERVAL: float = -1
VLLM_DISABLE_COMPILE_CACHE: bool = False
VLLM_REPLICATE_EMBED: bool = False
VLLM_USE_LAYERNAME: bool = True
Q_SCALE_CONSTANT: int = 200
K_SCALE_CONSTANT: int = 200
@@ -620,6 +621,9 @@ environment_variables: dict[str, Callable[[], Any]] = {
# Enable batch-invariant mode: deterministic results regardless of
# batch composition. Requires NVIDIA GPU with compute capability >= 9.0.
"VLLM_BATCH_INVARIANT": lambda: bool(int(os.getenv("VLLM_BATCH_INVARIANT", "0"))),
"VLLM_REPLICATE_EMBED": lambda: (
os.getenv("VLLM_REPLICATE_EMBED", "0").strip().lower() in ("1", "true")
),
# Use tensor descriptors for Q/K/V loads and output stores in the
# Triton unified-attention kernel. Enables HW 2D block reads on
# Intel XPU; the non-TD branch is dead-code-eliminated at Triton